Posts in category: Amazon AWS

AutoML allows you to derive rapid, general insights from your data right at the beginning of a machine learning (ML) project lifecycle. Understanding up front which preprocessing techniques and algorithm types provide best results reduces the time to develop, train, and deploy the right model. It plays a crucial role in every model’s development process and allows data scientists to focus on the most promising ML techniques. Additionally, AutoML provides a baseline model performance that can serve as a reference point for the data science team.

An AutoML tool applies a combination of different algorithms and various preprocessing techniques to your data. For example, it can scale the data, perform univariate feature selection, conduct PCA at different variance threshold levels, and apply clustering. Such preprocessing techniques could be applied individually or be combined in a pipeline. Subsequently, an AutoML tool would train different model types, such as Linear Regression, Elastic-Net, or Random Forest, on different versions of your preprocessed dataset and perform hyperparameter optimization (HPO). Amazon SageMaker Autopilot eliminates the heavy lifting of building ML models. After providing the dataset, SageMaker Autopilot automatically explores different solutions to find the best model. But what if you want to deploy your tailored version of an AutoML workflow?

This post shows how to create a custom-made AutoML workflow on Amazon SageMaker using Amazon SageMaker Automatic Model Tuning with sample code available in a GitHub repo.

Solution overview

For this use case, let’s assume you are part of a data science team that develops models in a specialized domain. You have developed a set of custom preprocessing techniques and selected a number of algorithms that you typically expect to work well with your ML problem. When working on new ML use cases, you would like first to perform an AutoML run using your preprocessing techniques and algorithms to narrow down the scope of potential solutions.

For this example, you don’t use a specialized dataset; instead, you work with the California Housing dataset that you will import from Amazon Simple Storage Service (Amazon S3). The focus is to demonstrate the technical implementation of the solution using SageMaker HPO, which later can be applied to any dataset and domain.

The following diagram presents the overall solution workflow.

Prerequisites

The following are prerequisites for completing the walkthrough in this post:

• An AWS account
• Familiarity with SageMaker concepts, such as an Estimator, training job, and HPO job
• Familiarity with the Amazon SageMaker Python SDK
• Python programming knowledge

Implement the solution

The full code is available in the GitHub repo.

The steps to implement the solution (as noted in the workflow diagram) are as follows:

1. Create a notebook instance and specify the following:
   1. For Notebook instance type, choose ml.t3.medium.
   2. For Elastic Inference, choose none.
   3. For Platform identifier, choose Amazon Linux 2, Jupyter Lab 3.
   4. For IAM role, choose the default AmazonSageMaker-ExecutionRole. If it doesn’t exist, create a new AWS Identity and Access Management (IAM) role and attach the AmazonSageMakerFullAccess IAM policy.

Note that you should create a minimally scoped execution role and policy in production.

2. Open the JupyterLab interface for your notebook instance and clone the GitHub repo.

You can do that by starting a new terminal session and running the git clone <REPO> command or by using the UI functionality, as shown in the following screenshot.

3. Open the automl.ipynb notebook file, select the conda_python3 kernel, and follow the instructions to trigger a set of HPO jobs.

To run the code without any changes, you need to increase the service quota for ml.m5.large for training job usage and Number of instances across all training jobs. AWS allows by default only 20 parallel SageMaker training jobs for both quotas. You need to request a quota increase to 30 for both. Both quota changes should typically be approved within a few minutes. Refer to Requesting a quota increase for more information.

If you don’t want to change the quota, you can simply modify the value of the MAX_PARALLEL_JOBS variable in the script (for example, to 5).

4. Each HPO job will complete a set of training job trials and indicate the model with optimal hyperparameters.
5. Analyze the results and deploy the best-performing model.

This solution will incur costs in your AWS account. The cost of this solution will depend on the number and duration of HPO training jobs. As these increase, so will the cost. You can reduce costs by limiting training time and configuring TuningJobCompletionCriteriaConfig according to the instructions discussed later in this post. For pricing information, refer to Amazon SageMaker Pricing.

In the following sections, we discuss the notebook in more detail with code examples and the steps to analyze the results and select the best model.

Initial setup

Let’s start with running the Imports & Setup section in the custom-automl.ipynb notebook. It installs and imports all the required dependencies, instantiates a SageMaker session and client, and sets the default Region and S3 bucket for storing data.

Data preparation

Download the California Housing dataset and prepare it by running the Download Data section of the notebook. The dataset is split into training and testing data frames and uploaded to the SageMaker session default S3 bucket.

The entire dataset has 20,640 records and 9 columns in total, including the target. The goal is to predict the median value of a house (medianHouseValue column). The following screenshot shows the top rows of the dataset.

Training script template

The AutoML workflow in this post is based on scikit-learn preprocessing pipelines and algorithms. The aim is to generate a large combination of different preprocessing pipelines and algorithms to find the best-performing setup. Let’s start with creating a generic training script, which is persisted locally on the notebook instance. In this script, there are two empty comment blocks: one for injecting hyperparameters and the other for the preprocessing-model pipeline object. They will be injected dynamically for each preprocessing model candidate. The purpose of having one generic script is to keep the implementation DRY (don’t repeat yourself).

#create base script
_script = """
import argparse
import joblib
import os
import numpy as np
import pandas as pd
from sklearn.metrics import mean_squared_error
from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.preprocessing import StandardScaler
from sklearn.decomposition import PCA
from sklearn.impute import SimpleImputer
from sklearn.cluster import KMeans
from sklearn.linear_model import ElasticNet
from sklearn.ensemble import RandomForestRegressor
from sklearn.ensemble import GradientBoostingRegressor
###########################
### Inference functions ###
###########################
def model_fn(model_dir):
clf = joblib.load(os.path.join(model_dir, "model.joblib"))
return clf
if __name__ == "__main__":
print("Extracting arguments")
parser = argparse.ArgumentParser()
# Hyperparameters
##### WILL BE INSERTED DYNAMICALLY #####
{}
############################
# Data, model, and output directories
parser.add_argument("--model-dir", type=str, default=os.environ.get("SM_MODEL_DIR"))
parser.add_argument("--train", type=str, default=os.environ.get("SM_CHANNEL_TRAIN"))
parser.add_argument("--test", type=str, default=os.environ.get("SM_CHANNEL_TEST"))
parser.add_argument("--train-file", type=str, default="train.parquet")
parser.add_argument("--test-file", type=str, default="test.parquet")
parser.add_argument("--features", type=str)
parser.add_argument("--target", type=str)
args, _ = parser.parse_known_args()
# Load and prepare data
train_df = pd.read_parquet(os.path.join(args.train, args.train_file))
test_df = pd.read_parquet(os.path.join(args.test, args.test_file))
X_train = train_df[args.features.split()]
X_test = test_df[args.features.split()]
y_train = train_df[args.target]
y_test = test_df[args.target]
# Train model
##### WILL BE INSERTED DYNAMICALLY #####
{}
{}
############################
pipeline = Pipeline([('preprocessor', preprocessor), ('model', model)])
pipeline.fit(X_train, y_train)
# Validate model and print metrics
rmse = mean_squared_error(y_test, pipeline.predict(X_test), squared=False)
print("RMSE: " + str(rmse))
# Persist model
path = os.path.join(args.model_dir, "model.joblib")
joblib.dump(pipeline, path)
"""
# write _script to file just to have it in hand
with open("script_draft.py", "w") as f:
print(_script, file=f)

Create preprocessing and model combinations

The preprocessors dictionary contains a specification of preprocessing techniques applied to all input features of the model. Each recipe is defined using a Pipeline or a FeatureUnion object from scikit-learn, which chains together individual data transformations and stack them together. For example, mean-imp-scale is a simple recipe that ensures that missing values are imputed using mean values of respective columns and that all features are scaled using the StandardScaler. In contrast, the mean-imp-scale-pca recipe chains together a few more operations:

1. Impute missing values in columns with its mean.
2. Apply feature scaling using mean and standard deviation.
3. Calculate PCA on top of the input data at a specified variance threshold value and merge it together with the imputed and scaled input features.

In this post, all input features are numeric. If you have more data types in your input dataset, you should specify a more complicated pipeline where different preprocessing branches are applied to different feature type sets.

preprocessors = {
    "mean-imp-scale": "preprocessor = Pipeline([('imputer', SimpleImputer(strategy='mean')), ('scaler', StandardScaler())])n",

    "mean-imp-scale-knn": "preprocessor = FeatureUnion([('base-features', Pipeline([('imputer', SimpleImputer(strategy='mean')), ('scaler', StandardScaler())])), ('knn', Pipeline([('imputer', SimpleImputer(strategy='mean')), ('scaler', StandardScaler()), ('knn', KMeans(n_clusters=10))]))])n",

    "mean-imp-scale-pca": "preprocessor = FeatureUnion([('base-features', Pipeline([('imputer', SimpleImputer(strategy='mean')), ('scaler', StandardScaler())])), ('pca', Pipeline([('imputer', SimpleImputer(strategy='mean')), ('scaler', StandardScaler()), ('pca', PCA(n_components=0.9))]))])n"   
}

The models dictionary contains specifications of different algorithms that you fit the dataset to. Every model type comes with the following specification in the dictionary:

• script_output – Points to the location of the training script used by the estimator. This field is filled dynamically when the models dictionary is combined with the preprocessors dictionary.
• insertions – Defines code that will be inserted into the script_draft.py and subsequently saved under script_output. The key “preprocessor” is intentionally left blank because this location is filled with one of the preprocessors in order to create multiple model-preprocessor combinations.
• hyperparameters – A set of hyperparameters that are optimized by the HPO job.
• include_cls_metadata – More configuration details required by the SageMaker Tuner class.

A full example of the models dictionary is available in the GitHub repository.

models = {
    "rf": {
        "script_output": None,
        "insertions": {
            # Arguments
            "arguments" : 
            "parser.add_argument('--n_estimators', type=int, default=100)n"+
            "    parser.add_argument('--max_depth', type=int, default=None)n"+            
            "    parser.add_argument('--min_samples_leaf', type=int, default=1)n"+
            "    parser.add_argument('--min_samples_split', type=int, default=2)n"+            
            "    parser.add_argument('--max_features', type=str, default='auto')n",
            # Model call
            "preprocessor": None,
            "model_call" : "model = RandomForestRegressor(n_estimators=args.n_estimators,max_depth=args.max_depth,min_samples_leaf=args.min_samples_leaf,min_samples_split=args.min_samples_split,max_features=args.max_features)n"
        },
        "hyperparameters": {
            "n_estimators": IntegerParameter(100, 2000, "Linear"),
            "max_depth": IntegerParameter(1, 100, "Logarithmic"),
            "min_samples_leaf": IntegerParameter(1, 6, "Linear"),
            "min_samples_split": IntegerParameter(2, 20, "Linear"),
            "max_features": CategoricalParameter(["auto", "sqrt", "log2"]),
        },
        "include_cls_metadata": False,
    }
}

Next, let’s iterate through the preprocessors and models dictionaries and create all possible combinations. For example, if your preprocessors dictionary contains 10 recipes and you have 5 model definitions in the models dictionary, the newly created pipelines dictionary contains 50 preprocessor-model pipelines that are evaluated during HPO. Note that individual pipeline scripts are not created yet at this point. The next code block (cell 9) of the Jupyter notebook iterates through all preprocessor-model objects in the pipelines dictionary, inserts all relevant code pieces, and persists a pipeline-specific version of the script locally in the notebook. Those scripts are used in the next steps when creating individual estimators that you plug into the HPO job.

pipelines = {}
for model_name, model_spec in models.items():
    pipelines[model_name] = {}
    for preprocessor_name, preprocessor_spec in preprocessors.items():
        pipeline_name = f"{model_name}-{preprocessor_name}"
        pipelines[model_name][pipeline_name] = {}
        pipelines[model_name][pipeline_name]["insertions"] = {}
        pipelines[model_name][pipeline_name]["insertions"]["preprocessor"] = preprocessor_spec
        pipelines[model_name][pipeline_name]["hyperparameters"] = model_spec["hyperparameters"]
        pipelines[model_name][pipeline_name]["include_cls_metadata"] = model_spec["include_cls_metadata"]        
        pipelines[model_name][pipeline_name]["insertions"]["arguments"] = model_spec["insertions"]["arguments"]
        pipelines[model_name][pipeline_name]["insertions"]["model_call"] = model_spec["insertions"]["model_call"]
        pipelines[model_name][pipeline_name]["script_output"] = f"scripts/{model_name}/script-{pipeline_name}.py"

Define estimators

You can now work on defining SageMaker Estimators that the HPO job uses after scripts are ready. Let’s start with creating a wrapper class that defines some common properties for all estimators. It inherits from the SKLearn class and specifies the role, instance count, and type, as well as which columns are used by the script as features and the target.

class SKLearnBase(SKLearn):
    def __init__(
        self, 
        entry_point=".", # intentionally left blank, will be overwritten in the next function
        framework_version="1.2-1",
        role=sm_role,
        instance_count=1,
        instance_type="ml.c5.xlarge",
        hyperparameters={
           "features": "medianIncome housingMedianAge totalRooms totalBedrooms population households latitude longitude",
            "target": "medianHouseValue",
        },  
        **kwargs,
        ):
        super(SKLearnBase, self).__init__(
            entry_point=entry_point,
            framework_version=framework_version,
            role=role,
            instance_count=instance_count,
            instance_type=instance_type,
            hyperparameters=hyperparameters,
            **kwargs
        )

Let’s build the estimators dictionary by iterating through all scripts generated before and located in the scripts directory. You instantiate a new estimator using the SKLearnBase class, with a unique estimator name, and one of the scripts. Note that the estimators dictionary has two levels: the top level defines a pipeline_family. This is a logical grouping based on the type of models to evaluate and is equal to the length of the models dictionary. The second level contains individual preprocessor types combined with the given pipeline_family. This logical grouping is required when creating the HPO job.

estimators = {}
for pipeline_family in pipelines.keys():
    estimators[pipeline_family] = {}
    scripts = os.listdir(f"scripts/{pipeline_family}")
    for script in scripts:
        if script.endswith(".py"):
            estimator_name = script.split(".")[0].replace("_", "-").replace("script", "estimator")
            estimators[pipeline_family][estimator_name] = SKLearnBase(
                entry_point=f"scripts/{pipeline_family}/{script}",
                base_job_name=estimator_name,
            )

Define HPO tuner arguments

To optimize passing arguments into the HPO Tuner class, the HyperparameterTunerArgs data class is initialized with arguments required by the HPO class. It comes with a set of functions, which ensure HPO arguments are returned in a format expected when deploying multiple model definitions at once.

@dataclass
class HyperparameterTunerArgs:
    base_job_names: list[str]
    estimators: list[object]
    inputs: dict[str]
    objective_metric_name: str
    hyperparameter_ranges: list[dict]
    metric_definition: dict[str]
    include_cls_metadata: list[bool]

    def get_estimator_dict(self) -> dict:
        return {k:v for (k, v) in zip(self.base_job_names, self.estimators)}

    def get_inputs_dict(self) -> dict:
        return {k:v for (k, v) in zip(self.base_job_names, [self.inputs]*len(self.base_job_names))}

    def get_objective_metric_name_dict(self) -> dict:
        return {k:v for (k, v) in zip(self.base_job_names, [self.objective_metric_name]*len(self.base_job_names))}

    def get_hyperparameter_ranges_dict(self) -> dict:
        return {k:v for (k, v) in zip(self.base_job_names, self.hyperparameter_ranges)}

    def get_metric_definition_dict(self) -> dict:
        return {k:[v] for (k, v) in zip(self.base_job_names, [self.metric_definition]*len(self.base_job_names))}

    def get_include_cls_metadata_dict(self) -> dict:
        return {k:v for (k, v) in zip(self.base_job_names, self.include_cls_metadata)}

The next code block uses the previously introduced HyperparameterTunerArgs data class. You create another dictionary called hp_args and generate a set of input parameters specific to each estimator_family from the estimators dictionary. These arguments are used in the next step when initializing HPO jobs for each model family.

hp_args = {}
for estimator_family, estimators in estimators.items():
    hp_args[estimator_family] = HyperparameterTunerArgs(
        base_job_names=list(estimators.keys()),
        estimators=list(estimators.values()),
        inputs={"train": s3_data_train.uri, "test": s3_data_test.uri},
        objective_metric_name="RMSE",
        hyperparameter_ranges=[pipeline.get("hyperparameters") for pipeline in pipelines[estimator_family].values()],
        metric_definition={"Name": "RMSE", "Regex": "RMSE: ([0-9.]+).*$"},
        include_cls_metadata=[pipeline.get("include_cls_metadata") for pipeline in pipelines[estimator_family].values()],
    )

Create HPO tuner objects

In this step, you create individual tuners for every estimator_family. Why do you create three separate HPO jobs instead of launching just one across all estimators? The HyperparameterTuner class is restricted to 10 model definitions attached to it. Therefore, each HPO is responsible for finding the best-performing preprocessor for a given model family and tuning that model family’s hyperparameters.

The following are a few more points regarding the setup:

• The optimization strategy is Bayesian, which means that the HPO actively monitors the performance of all trials and navigates the optimization towards more promising hyperparameter combinations. Early stopping should be set to Off or Auto when working with a Bayesian strategy, which handles that logic itself.
• Each HPO job runs for a maximum of 100 jobs and runs 10 jobs in parallel. If you’re dealing with larger datasets, you might want to increase the total number of jobs.
• Additionally, you may want to use settings that control how long a job runs and how many jobs your HPO is triggering. One way to do that is to set the maximum runtime in seconds (for this post, we set it to 1 hour). Another is to use the recently released TuningJobCompletionCriteriaConfig. It offers a set of settings that monitor the progress of your jobs and decide whether it is likely that more jobs will improve the result. In this post, we set the maximum number of training jobs not improving to 20. That way, if the score isn’t improving (for example, from the fortieth trial), you won’t have to pay for the remaining trials until max_jobs is reached.

STRATEGY = "Bayesian"
OBJECTIVE_TYPE = "Minimize"
MAX_JOBS = 100
MAX_PARALLEL_JOBS = 10
MAX_RUNTIME_IN_SECONDS = 3600
EARLY_STOPPING_TYPE = "Off"
# RANDOM_SEED = 42 # uncomment if you require reproducibility across runs
TUNING_JOB_COMPLETION_CRITERIA_CONFIG = TuningJobCompletionCriteriaConfig(
    max_number_of_training_jobs_not_improving=20,
    )

tuners = {}
for estimator_family, hp in hp_args.items():
    tuners[estimator_family] = HyperparameterTuner.create(
        estimator_dict=hp.get_estimator_dict(),
        objective_metric_name_dict=hp.get_objective_metric_name_dict(),
        hyperparameter_ranges_dict=hp.get_hyperparameter_ranges_dict(),
        metric_definitions_dict=hp.get_metric_definition_dict(),
        strategy=STRATEGY,
        completion_criteria_config=TUNING_JOB_COMPLETION_CRITERIA_CONFIG,
        objective_type=OBJECTIVE_TYPE,
        max_jobs=MAX_JOBS,
        max_parallel_jobs=MAX_PARALLEL_JOBS,
        max_runtime_in_seconds=MAX_RUNTIME_IN_SECONDS,
        base_tuning_job_name=f"custom-automl-{estimator_family}",
        early_stopping_type=EARLY_STOPPING_TYPE, # early stopping of training jobs is not currently supported when multiple training job definitions are used
        # random_seed=RANDOM_SEED,
    )

Now let’s iterate through the tuners and hp_args dictionaries and trigger all HPO jobs in SageMaker. Note the usage of the wait argument set to False, which means that the kernel won’t wait until the results are complete and you can trigger all jobs at once.

It’s likely that not all training jobs will complete and some of them might be stopped by the HPO job. The reason for this is the TuningJobCompletionCriteriaConfig—the optimization finishes if any of the specified criteria is met. In this case, when the optimization criteria isn’t improving for 20 consecutive jobs.

for tuner, hpo in zip(tuners.values(), hp_args.values()):
    tuner.fit(
        inputs=hpo.get_inputs_dict(),
        include_cls_metadata=hpo.get_include_cls_metadata_dict(),
        wait=False,
        )

Analyze results

Cell 15 of the notebook checks if all HPO jobs are complete and combines all results in the form of a pandas data frame for further analysis. Before analyzing the results in detail, let’s take a high-level look at the SageMaker console.

At the top of the Hyperparameter tuning jobs page, you can see your three launched HPO jobs. All of them finished early and didn’t perform all 100 training jobs. In the following screenshot, you can see that the Elastic-Net model family completed the highest number of trials, whereas others didn’t need so many training jobs to find the best result.

You can open the HPO job to access more details, such as individual training jobs, job configuration, and the best training job’s information and performance.

Let’s produce a visualization based on the results to get more insights of the AutoML workflow performance across all model families.

From the following graph, you can conclude that the Elastic-Net model’s performance was oscillating between 70,000 and 80,000 RMSE and eventually stalled, as the algorithm wasn’t able to improve its performance despite trying various preprocessing techniques and hyperparameter values. It also seems that RandomForest performance varied a lot depending on the hyperparameter set explored by HPO, but despite many trials it couldn’t go below the 50,000 RMSE error. GradientBoosting achieved the best performance already from the start going below 50,000 RMSE. HPO tried to improve that result further but wasn’t able to achieve better performance across other hyperparameter combinations. A general conclusion for all HPO jobs is that not so many jobs were required to find the best performing set of hyperparameters for each algorithm. To further improve the result, you would need to experiment with creating more features and performing additional feature engineering.

You can also examine a more detailed view of the model-preprocessor combination to draw conclusions about the most promising combinations.

Select the best model and deploy it

The following code snippet selects the best model based on the lowest achieved objective value. You can then deploy the model as a SageMaker endpoint.

df_best_job = df_tuner_results.loc[df_tuner_results["FinalObjectiveValue"] == df_tuner_results["FinalObjectiveValue"].min()]
df_best_job
BEST_MODEL_FAMILY = df_best_job["TrainingJobFamily"].values[0]

tuners.get(BEST_MODEL_FAMILY).best_training_job()

tuners.get(BEST_MODEL_FAMILY).best_estimator()

predictor = tuners.get(BEST_MODEL_FAMILY).deploy(
    initial_instance_count=1,
    instance_type="ml.c4.large",
    endpoint_name=f"custom-automl-endpoint-{BEST_MODEL_FAMILY}",
)

Clean up

To prevent unwanted charges to your AWS account, we recommend deleting the AWS resources that you used in this post:

1. On the Amazon S3 console, empty the data from the S3 bucket where the training data was stored.

2. On the SageMaker console, stop the notebook instance.

3. Delete the model endpoint if you deployed it. Endpoints should be deleted when no longer in use, because they’re billed by time deployed.

sm_client.delete_endpoint(EndpointName=predictor.endpoint)

Conclusion

In this post, we showcased how to create a custom HPO job in SageMaker using a custom selection of algorithms and preprocessing techniques. In particular, this example demonstrates how to automate the process of generating many training scripts and how to use Python programming structures for efficient deployment of multiple parallel optimization jobs. We hope this solution will form the scaffolding of any custom model tuning jobs you will deploy using SageMaker to achieve higher performance and speed up of your ML workflows.

Check out the following resources to further deepen your knowledge of how to use SageMaker HPO:

• Best Practices for Hyperparameter Tuning
• Amazon SageMaker Automatic Model Tuning now supports three new completion criteria for hyperparameter optimization
• Using Scikit-learn with the SageMaker Python SDK
• Develop, Train, Optimize and Deploy Scikit-Learn Random Forest

About the Authors

Konrad Semsch is a Senior ML Solutions Architect at Amazon Web Services Data Lab Team. He helps customers use machine learning to solve their business challenges with AWS. He enjoys inventing and simplifying to enable customers with simple and pragmatic solutions for their AI/ML projects. He is most passionate about MlOps and traditional data science. Outside of work, he is a big fan of windsurfing and kitesurfing.

Tuna Ersoy is a Senior Solutions Architect at AWS. Her primary focus is helping Public Sector customers adopt cloud technologies for their workloads. She has a background in application development, enterprise architecture, and contact center technologies. Her interests include serverless architectures and AI/ML.

Read More

Llama 2 stands at the forefront of AI innovation, embodying an advanced auto-regressive language model developed on a sophisticated transformer foundation. It’s tailored to address a multitude of applications in both the commercial and research domains with English as the primary linguistic concentration. Its model parameters scale from an impressive 7 billion to a remarkable 70 billion. Llama 2 demonstrates the potential of large language models (LLMs) through its refined abilities and precisely tuned performance.

Diving deeper into Llama 2’s architecture, Meta reveals that the model’s fine-tuning melds supervised fine-tuning (SFT) with reinforcement learning aided by human feedback (RLHF). This combination prioritizes alignment with human-centric norms, striking a balance between efficiency and safety. Built upon a vast reservoir of 2 trillion tokens, Llama 2 provides both pre-trained models for diverse natural language generation and the specialized Llama-2-Chat variant for chat assistant roles. Regardless of a developer’s choice between the basic or the advanced model, Meta’s responsible use guide is an invaluable resource for model enhancement and customization.

For those interested in creating interactive applications, Llama 2 Chat is a good starting point. This conversational model allows for building customized chatbots and assistants. To make it even more accessible, you can deploy Llama-2-Chat models with ease through Amazon SageMaker JumpStart. An offering from Amazon SageMaker, SageMaker JumpStart provides a straightforward way to deploy Llama-2 model variants directly through Amazon SageMaker Studio notebooks. This enables developers to focus on their application logic while benefiting from SageMaker tools for scalable AI model training and hosting. SageMaker JumpStart also provides effortless access to the extensive SageMaker library of algorithms and pre-trained models.

In this post, we explore best practices for prompting the Llama 2 Chat LLM. We highlight key prompt design approaches and methodologies by providing practical examples.

Prerequisites

To try out the examples and recommended best practices for Llama 2 Chat on SageMaker JumpStart, you need the following prerequisites:

• An AWS account that will contain all your AWS resources.
• An AWS Identity and Access Management (IAM) role to access SageMaker. To learn more about how IAM works with SageMaker, refer to Identity and Access Management for Amazon SageMaker.
• Access to SageMaker Studio or a SageMaker notebook instance or an interactive development environment (IDE) such as PyCharm or Visual Studio Code. We recommend using SageMaker Studio notebooks for straightforward deployment and inference.
• The GitHub repository cloned in order to use the accompanying notebook.
• An instance of Llama 2 Chat model deployed on SageMaker using SageMaker JumpStart. To learn more, refer to Llama 2 foundation models from Meta are now available in Amazon SageMaker JumpStart. The accompanying notebook also contains code to deploy the model.

Prompting techniques

Prompting, in the context of language models and artificial intelligence, refers to the practice of providing a model with a specific input or cue to elicit a desired response. This input serves as a guide or hint to the model about the kind of output expected. Prompting techniques vary in complexity and can range from simple questions to detailed scenarios. Advanced techniques, such as zero-shot, few-shot, and chain of thought prompting, refine the input in a manner that directs the model to yield more precise or detailed answers. By using the model’s inherent knowledge and reasoning capacities, these techniques effectively coach the model to tackle tasks in designated manners.

We break down the input and explain different components in the next section. We start by sharing some examples of what different prompt techniques look like. The examples are always shown in two code blocks. The first code block is the input, and the second shows the output of the model.

Zero-shot prompting

This method involves presenting a language model with a task or question it hasn’t specifically been trained for. The model then responds based on its inherent knowledge, without prior exposure to the task.

%%time

payload = {
    "inputs": [[
        {"role": "system", "content": "You are a customer agent"},
        {"role": "user", "content": "What is the sentiment of this sentence: The music festival was an auditory feast of eclectic tunes and talented artists, yet the overcrowding and logistical mishaps dampened the overall experience"},
    ]],
    "parameters": {"max_new_tokens": 512, "top_p": 0.9, "temperature": 0.6}
}
response = predictor.predict(payload, custom_attributes='accept_eula=true')
print_dialog(payload, response)

System: You are a customer agent

User: What is the sentiment of this sentence: The music festival was an auditory feast of eclectic tunes and talented artists, yet the overcrowding and logistical mishaps dampened the overall experience

======================================================

Assistant:  The sentiment of the sentence is neutral. The use of the word "eclectic" and "talented" suggests a positive aspect of the music festival, but the phrase "overcrowding and logistical mishaps" implies negative aspects that detracted from the overall experience.
======================================================

CPU times: user 4.24 ms, sys: 389 µs, total: 4.63 ms
Wall time: 1.19 s

Few-shot prompting

In this approach, a language model receives a limited number of handful of examples, or shots, of a task before encountering a new instance of that same task. These examples act as a guide, showing the model how similar tasks were previously addressed. Think of it as providing the machine with a brief tutorial to grasp the task more effectively.

%%time

payload = {
    "inputs": [[
        {"role": "system", "content": "You are a customer agent"},
        {"role": "user", "content": f"""
                                    nnExample 1
                                    nSentence: Though the sun set with a brilliant display of colors, casting a warm glow over the serene beach, it was the bitter news I received earlier that clouded my emotions, making it impossible to truly appreciate nature's beauty.
                                    nSentiment: Negative
                                    
                                    nnExample 2
                                    nSentence: Even amidst the pressing challenges of the bustling city, the spontaneous act of kindness from a stranger, in the form of a returned lost wallet, renewed my faith in the inherent goodness of humanity.
                                    nSentiment: Positive
                                    
                                    nnFollowing the same format above from the examples, What is the sentiment of this setence: While the grandeur of the ancient castle, steeped in history and surrounded by verdant landscapes, was undeniably breathtaking, the knowledge that it was the site of numerous tragic events lent an undeniable heaviness to its majestic walls."""},
    ]],
    "parameters": {"max_new_tokens": 512, "top_p": 0.9, "temperature": 0.6}
}
response = predictor.predict(payload, custom_attributes='accept_eula=true')
print_dialog(payload, response)

System: You are a customer agent

User: 
                                    
Example 1
                                    
Sentence: Though the sun set with a brilliant display of colors, casting a warm glow over the serene beach, it was the bitter news I received earlier that clouded my emotions, making it impossible to truly appreciate nature's beauty.
                                    
Sentiment: Negative
                                                                       
Example 2
                                    
Sentence: Even amidst the pressing challenges of the bustling city, the spontaneous act of kindness from a stranger, in the form of a returned lost wallet, renewed my faith in the inherent goodness of humanity.
                                    
Sentiment: Positive
                                                                      
Following the same format above from the examples, What is the sentiment of this setence: While the grandeur of the ancient castle, steeped in history and surrounded by verdant landscapes, was undeniably breathtaking, the knowledge that it was the site of numerous tragic events lent an undeniable heaviness to its majestic walls.

======================================================

Assistant:  Sure! Here's the sentiment analysis of the given sentence:

Sentence: While the grandeur of the ancient castle, steeped in history and surrounded by verdant landscapes, was undeniably breathtaking, the knowledge that it was the site of numerous tragic events lent an undeniable heaviness to its majestic walls.

Sentiment: Mixed (both positive and negative)
======================================================

CPU times: user 3.72 ms, sys: 773 µs, total: 4.49 ms
Wall time: 1.85 s

Chain of thought prompting

This approach augments the reasoning capabilities of LLMs in intricate tasks. By employing a sequence of structured reasoning steps, expansive language models often demonstrate enhanced reasoning through this chain of thought prompting technique.

%%time

payload = {
    "inputs": [[
        {"role": "system", "content": "You are a pizza professional"},
        {"role": "user", "content": f"""
        You have a pizza that was cut into 8 equal slices. You ate 3 slices, and your friend ate 2 slices. Here's how we can figure out how many slices are left:

            1. Start with the total number of slices.
            2. Subtract the number of slices you ate.
            3. Then subtract the number of slices your friend ate.
            4. The result is the number of slices remaining.

            So, let's calculate:

        """},
    ]],
    "parameters": {"max_new_tokens": 512, "top_p": 0.9, "temperature": 0.6}
}
response = predictor.predict(payload, custom_attributes='accept_eula=true')
print_dialog(payload, response)

System: You are a pizza professional

User: 
        You have a pizza that was cut into 8 equal slices. You ate 3 slices, and your friend ate 2 slices. Here's how we can figure out how many slices are left:

            1. Start with the total number of slices.
            2. Subtract the number of slices you ate.
            3. Then subtract the number of slices your friend ate.
            4. The result is the number of slices remaining.

            So, let's calculate:
 
======================================================

Assistant:  Oh man, I love pizza! Alright, let's get started on this math problem. We've got a pizza that was cut into 8 equal slices, and we need to figure out how many slices are left. Here's the plan:

Step 1: Start with the total number of slices. That's 8 slices.

Step 2: Subtract the number of slices I ate. I ate 3 slices, so let's subtract 3 from 8.

8 - 3 = 5

Step 3: Subtract the number of slices my friend ate. My friend ate 2 slices, so let's subtract 2 from 5.

5 - 2 = 3

Step 4: The result is the number of slices remaining. In this case, there are 3 slices remaining!

So, there you have it! There are 3 slices of pizza left. I hope my friend and I can agree on how to divide them up! 🍕👍
======================================================

CPU times: user 11.8 ms, sys: 1.72 ms, total: 13.5 ms
Wall time: 4.7 s

In the preceding example, Llama 2 Chat was able to assume the persona of a professional that has domain knowledge and was able to demonstrate the reasoning in getting to a conclusion.

Llama 2 Chat inference parameters

Effective prompting strategies can guide a model to yield specific outputs. For those seeking a finer level of control over these outputs, Llama 2 Chat introduces a comprehensive set of inference parameters:

• max_new_tokens – Defines the length of the model’s output. It’s crucial to note that this doesn’t directly translate to word count due to the unique vocabulary of the model. A single token might represent more than one English word.
• temperature – Affects the randomness of the output. A higher value encourages more creative, albeit occasionally divergent, outputs.
• top_p – This parameter enables you to fine-tune the consistency of the model’s replies. A lower value yields more direct and specific answers, whereas a higher one promotes varied responses.

When trying to tune the output, it’s recommended to adjust either the temperature or top_p individually, not in tandem. Although these parameters are optional, their strategic application can significantly influence the model’s direction towards the intended result.

Introduction to system prompts

Llama 2 Chat uses a transformative feature called system prompts. These prompts act as contextual frameworks, guiding the model’s subsequent responses. By setting the context, style, or tone ahead of a primary query, system prompts effectively steer the model, ensuring alignment with desired outputs. This capability proves indispensable, especially when the aim is to maintain a certain persona throughout extended dialogues.

In our previous examples, the following part of the input sets the system context:

{"role": "system", "content": "You are a pizza professional"},

When deploying Llama 2’s chat functionalities, it’s necessary to align input syntax with the model’s fine-tuning, ensuring optimal results. It’s structured as follows:

• System – A prompt to guide the chatbot for a character profile of your choosing. It could be tailored as per your preference (like “you are a minion”). If you omit this, the model will no longer have a default generic profile. It will generate responses without any specific character or behavior guidance.

  {"role": "system", "content": "You are a minion"}

• User – A message from a user.

  {"role": "user", "content": "What do you think of bananas?"}

• Assistant – The output result from the model.

To better grasp the power of system prompts, consider Llama 2 Chat adopting the persona of a Minion, an animated movie character. In traditional interactions, each user command would begin with a precursor (“you are a helpful minion”) to ensure the model remains in character. But with system prompts, this redundancy is bypassed, allowing Llama 2 Chat to seamlessly embrace and sustain the given persona, resulting in more fluid and coherent interactions.

Practical demonstration

Let’s start with the following code to demonstrate:

%%time

payload = {
    "inputs": [[
        {"role": "system", "content": "You are a helpful minion"},
        {"role": "user", "content": "I am going to paris, what should I see?"},
    ]],
    "parameters": {"max_new_tokens": 512, "top_p": 0.9, "temperature": 0.6}
}
response = predictor.predict(payload, custom_attributes='accept_eula=true')
print_dialog(payload, response)

System: You are a helpful minion

User: I am going to paris, what should I see?

======================================================

Assistant:  Oooh la la! *giggle* Paris, the city of love and romance! *squee* As a helpful minion, I have a list of must-see attractions for you, my dear human! *bats eyelashes*

1. The Eiffel Tower: *swoon* The iconic iron lady of Paris! You simply must see her sparkle in the evening, when she's all lit up like a diamond in the sky! *heart eyes*
2. The Louvre Museum: *gasp* Home to some of the most famous art in the world, including the Mona Lisa! *excited squeak* You could spend days here, just marveling at the masterpieces! *bouncy bouncy*
3. Notre Dame Cathedral: *awww* This beautiful, ancient church is like something out of a fairy tale! *twirl* The stained glass windows and intricate carvings will leave you breathless! *sigh*
4. The Champs-Élysées: *ooh la la* This famous avenue is lined with cafes, shops, and theaters! *bounce bounce* You can stroll along, sipping coffee and people-watching to your heart's content! *wink wink*
5. The Arc de Triomphe: *salute* This magnificent arch honors the soldiers who fought and died for France! *proudly* The views from the top are simply stunning! *look around*
6. Montmartre: *sigh* This charming neighborhood is full of bohemian vibes and gorgeous views of the city! *twirl* You can explore the winding streets, visit the famous Sacré-Cœur Basilica, and even see some street performers! *bats eyelashes*
7. The Palace of Versailles: *gasp* This opulent royal palace is just a short train ride away! *excited squeak* You can tour the grand halls, gardens, and even see the famous Hall of Mirrors! *twirl*

Ooh la la! I hope you have a simply marvelous time in Paris, my dear human! *curtsy* Don't
======================================================

CPU times: user 13.5 ms, sys: 2.24 ms, total: 15.8 ms
Wall time: 9.55 s

This interaction exemplifies the model’s proficiency in adopting the helpful Minion persona, offering users a consistent and engaging experience while reducing input repetition.

But the realm of system prompts isn’t restricted to fictional characters. Its versatility shines in more serious contexts as well. Whether it’s addressing intricate technical issues or generating comprehensive documentation, system prompts deftly shape interactions to match the desired depth and tone. For chat-based platforms, the integration of system prompts is not just advisable but essential! Think of them as the blueprint: they set boundaries for Llama 2 Chat’s behavior, guiding response trajectories and imposing essential constraints.

Professional context demonstration

In this next example, we ask the model to assume a more professional demeanor:

%%time

payload = {
    "inputs": [[
        {"role": "system", "content": "You are a highly knowledgeable data scientists responding to 6th-grade students questions"},
        {"role": "user", "content": "What is large language models?"},
    ]],
    "parameters": {"max_new_tokens": 512, "top_p": 0.9, "temperature": 0.6}
}
response = predictor.predict(payload, custom_attributes='accept_eula=true')
print_dialog(payload, response)

System: You are a highly knowledgeable data scientists responding to 6th-grade students questions

User: What is large language models?

======================================================
 
Assistant:  Wow, you must be really curious about language models! adjusts glasses Large language models are like superheroes for words! 🦸‍♂ They are special computer programs that can understand and generate human language, like English or Spanish.
 
Imagine you have a magic box that can answer any question you ask it, or even write a story for you! That's kind of like what large language models can do. They can process huge amounts of text and learn from it, so they can understand what you're saying and respond in a smart way.
 
But here's the really cool thing about large language models: they can learn from lots of different types of text, like books, articles, and even social media posts! 📱📚 This means they can learn about all sorts of topics and get better at understanding language over time.
 
So, why are large language models important? Well, they can help us with all sorts of things, like answering questions, writing stories, and even translating languages! 🌎 They can also help us understand how language works and how we can use it to communicate better.
 
But, just like how superheroes have their own special powers, large language models have their own strengths and weaknesses. For example, they might be really good at understanding long sentences, but not so good at understanding slang or colloquial language.
 
So, there you have it! Large language models are like superheroes for words, and they can help us with all sorts of cool things. Do you have any more questions about them? 🤔
 
======================================================
 
CPU times: user 960 µs, sys: 4.32 ms, total: 5.28 ms
Wall time: 6.68 s

The accuracy and controllability of the outputs are vital. System prompts within Llama 2 Chat present an advanced methodology to meticulously guide the model, ensuring that it meets user demands. Let’s delve deeper with two illustrative use cases:

• Scenario 1 – Envisaging the model as a knowledge English professor, a user seeks an in-depth analysis from a given synopsis. The model’s output mirrors a conversation one might have with an English scholar, enveloped amidst classical literature:

%%time

payload = {
    "inputs": [[
        {"role": "system", "content": "You are a profoundly knowledgeable English professor"},
        {"role": "user", "content": f"""Extract the main character's name, their main objective, the main obstacle they face, and the setting of the story from the following synopsis: "In a dystopian city of the future, Eleanor seeks to overthrow the oppressive regime that controls every aspect of citizens' lives. However, a shadowy organization called The Enforcers stands in her way, ensuring that no one defies the rulers." """},
    ]],
    "parameters": {"max_new_tokens": 512, "top_p": 0.9, "temperature": 0.6}
}
response = predictor.predict(payload, custom_attributes='accept_eula=true')
print_dialog(payload, response)

System: You are a profoundly knowledgeable English professor

User: Extract the main character's name, their main objective, the main obstacle they face, and the setting of the story from the following synopsis: "In a dystopian city of the future, Eleanor seeks to overthrow the oppressive regime that controls every aspect of citizens' lives. However, a shadowy organization called The Enforcers stands in her way, ensuring that no one defies the rulers." 

======================================================


Assistant:  Ah, a most intriguing synopsis! Let me extract the main elements for you:

1. Main character's name: Eleanor
2. Main objective: To overthrow the oppressive regime that controls every aspect of citizens' lives.
3. Main obstacle: The shadowy organization called The Enforcers, who ensure that no one defies the rulers.
4. Setting: A dystopian city of the future.

Ah, a most compelling story! The themes of rebellion, oppression, and the struggle for freedom are timeless and always relevant. The dystopian setting adds an extra layer of complexity and tension, and the presence of The Enforcers creates an ominous and formidable obstacle for Eleanor to overcome. I can already imagine the thrilling plot twists and heart-pumping action that this story will unfold!
======================================================

CPU times: user 4.42 ms, sys: 154 µs, total: 4.58 ms
Wall time: 3.73 s

• Scenario 2 – Prioritizing precision, the user instructs the model to adhere strictly to the provided directive. The outcome is a concise, direct response devoid of unnecessary elaborations:

%%time

payload = {
    "inputs": [[
        {"role": "system", "content": "You are a profoundly knowledgeable English professor who follows every instruction meticulously."},
        {"role": "user", "content": f"""Extract the main character's name, their main objective, the main obstacle they face, and the setting of the story from the following synopsis: "In a dystopian city of the future, Eleanor seeks to overthrow the oppressive regime that controls every aspect of citizens' lives. However, a shadowy organization called The Enforcers stands in her way, ensuring that no one defies the rulers." """},
    ]],
    "parameters": {"max_new_tokens": 512, "top_p": 0.9, "temperature": 0.6}
}
response = predictor.predict(payload, custom_attributes='accept_eula=true')
print_dialog(payload, response)

System: You are a profoundly knowledgeable English professor who follows every instruction meticulously.

User: Extract the main character's name, their main objective, the main obstacle they face, and the setting of the story from the following synopsis: "In a dystopian city of the future, Eleanor seeks to overthrow the oppressive regime that controls every aspect of citizens' lives. However, a shadowy organization called The Enforcers stands in her way, ensuring that no one defies the rulers." 

======================================================

Assistant:  Ah, an intriguing synopsis! Let me extract the main character's name, main objective, main obstacle, and setting for you:

1. Main character's name: Eleanor
2. Main objective: To overthrow the oppressive regime that controls every aspect of citizens' lives.
3. Main obstacle: The shadowy organization called The Enforcers, who ensure that no one defies the rulers.
4. Setting: A dystopian city of the future.

Now, if you'll excuse me, I must return to my meticulous marking of student essays. The rigors of academia demand nothing less!
======================================================

CPU times: user 2.16 ms, sys: 1.74 ms, total: 3.89 ms
Wall time: 2.84 s

The distinction between these scenarios underlines the adaptive nature of system prompts. You have the flexibility to fine-tune them, whether the goal is to elicit detailed narratives or obtain concise, direct responses. Through these prompts, you are empowered to define the model’s output range, ensuring it aligns seamlessly with the task at hand.

Text summarization and in-context question answering

System prompts can also be straightforward and enforce context to answer questions. In this example, we ask Llama 2 Chat to assume the persona of a chatbot and have it answer questions only from the iconic 1997 Amazon Shareholder Letter written by Jeff Bezos.

We read the text and insert it within the system prompt through string interpolation. The complete example is shown in the accompanying notebook.

system_prompt = f"""
You are an intelligent chatbot. Answer the questions only using the following context:

{Amazon_Shareholder_Letter_1997}

Here are some rules you always follow:

- Generate human readable output, avoid creating output with gibberish text.
- Generate only the requested output, don't include any other language before or after the requested output.
- Never say thank you, that you are happy to help, that you are an AI agent, etc. Just answer directly.
- Generate professional language typically used in business documents in North America.
- Never generate offensive or foul language.
"""

user_prompt = "Give me the summary of the shareholder letter"

%%time

payload = {
    "inputs": [[
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_prompt},
    ]],
    "parameters": {"max_new_tokens": 512, "top_p": 0.9, "temperature": 0.6}
}
response = predictor.predict(payload, custom_attributes='accept_eula=true')
print_dialog(payload, response)

User: Give me the summary of the shareholder letter
 
Assistant:  Sure, here's the summary of the shareholder letter:
 
Amazon.com's Founder and CEO, Jeffrey P. Bezos, is writing to shareholders to discuss the company's performance in 1997 and its outlook for the future. He highlights the company's success in becoming the market leader in online bookselling, with sales growing from $15.7 million in 1996 to $147.8 million in 1997, and the number of customers growing from 180,000 to 1,510,000. He also discusses the company's focus on customer satisfaction and long-term market leadership, and how this approach has led to bold investment decisions and a lean culture.
 
Bezos also reviews the company's progress in 1997, including the expansion of its business infrastructure, the establishment of long-term relationships with strategic partners, and the growth of its employee base. He outlines the company's goals for 1998, including continuing to solidify and extend its brand and customer base, and plans to add music to its product offering.
 
Overall, the letter emphasizes Amazon.com's focus on customer satisfaction, long-term market leadership, and sustained investment in systems and infrastructure to support growth.
 
======================================================
 
CPU times: user 17.9 ms, sys: 1.8 ms, total: 19.7 ms
Wall time: 8.08 s

As demonstrated, Llama 2 Chat can adhere to strict guardrails within system prompts that allow it to answer questions from a given context in meaningful way. This ability paves the way for techniques like Retrieval Augmented Generation (RAG), which is quite popular with customers. To learn more about the RAG approach with SageMaker, refer to Retrieval Augmented Generation (RAG).

Conclusion

Deploying Llama 2 Chat to achieve strong performance requires both technical expertise and strategic insight into its design. To fully take advantage of the model’s extensive abilities, you must understand and apply creative prompting techniques and adjust inference parameters. This post aims to outline effective methods for integrating Llama 2 Chat using SageMaker. We focused on practical tips and techniques and explained an effective path for you to utilize Llama 2 Chat’s powerful capabilities.

The following are key takeaways:

• Dynamic control with ambience – The temperature controls within Llama 2 Chat serve a pivotal role far beyond simple adjustments. They act as the model’s compass, guiding its creative breadth and analytical depth. Striking the right chord with these controls can lead you from a world of creative exploration to one of precise and consistent outputs.
• Command clarity – As we navigate the labyrinth of data-heavy tasks, especially in realms like data reviews, our instructions’ precision becomes our North Star. Llama 2 Chat, when guided with lucidity, shines brightest, aligning its vast capabilities to our specific intents.
• Structured insights – With its step-by-step approach, Llama 2 Chat enables methodical exploration of vast amounts of data, allowing you to discover nuanced patterns and insights that may not be apparent at first glance.

Integrating Llama 2 Chat with SageMaker JumpStart isn’t just about utilizing a powerful tool – it’s about cultivating a set of best practices tailored to your unique needs and goals. Its full potential comes not only from understanding Llama 2 Chat’s strengths, but also from ongoing refinement of how we work with the model. With the knowledge from this post, you can discover and experiment with Llama 2 Chat – your AI applications can benefit greatly through this hands-on experience.

Resources

• Llama 2 foundation models from Meta are now available in Amazon SageMaker JumpStart
• Fine-tune Llama 2 for text generation on Amazon SageMaker JumpStart
• Improve throughput performance of Llama 2 models using Amazon SageMaker

About the authors

Jin Tan Ruan is a Prototyping Developer within the AWS Industries Prototyping and Customer Engineering (PACE) team, specializing in NLP and generative AI. With a background in software development and nine AWS certifications, Jin brings a wealth of experience to assist AWS customers in materializing their AI/ML and generative AI visions using the AWS platform. He holds a master’s degree in Computer Science & Software Engineering from the University of Syracuse. Outside of work, Jin enjoys playing video games and immersing himself in the thrilling world of horror movies. You can find Jin on Linkedln. Let’s connect!

Dr. Farooq Sabir is a Senior Artificial Intelligence and Machine Learning Specialist Solutions Architect at AWS. He holds PhD and MS degrees in Electrical Engineering from the University of Texas at Austin and an MS in Computer Science from Georgia Institute of Technology. He has over 15 years of work experience and also likes to teach and mentor college students. At AWS, he helps customers formulate and solve their business problems in data science, machine learning, computer vision, artificial intelligence, numerical optimization, and related domains. Based in Dallas, Texas, he and his family love to travel and go on long road trips.

Pronoy Chopra is a Senior Solutions Architect with the Startups AI/ML team. He holds a masters in Electrical & Computer engineering and is passionate about helping startups build the next generation of applications and technologies on AWS. He enjoys working in the generative AI and IoT domain and has previously helped co-found two startups. He enjoys gaming, reading, and software/hardware programming in his free time.

Read More

An established financial services firm with over 140 years in business, Principal is a global investment management leader and serves more than 62 million customers around the world. Principal is conducting enterprise-scale near-real-time analytics to deliver a seamless and hyper-personalized omnichannel customer experience on their mission to make financial security accessible for all. They are processing data across channels, including recorded contact center interactions, emails, chat and other digital channels.

In this post, we demonstrate how data aggregated within the AWS CCI Post Call Analytics solution allowed Principal to gain visibility into their contact center interactions, better understand the customer journey, and improve the overall experience between contact channels while also maintaining data integrity and security.

Solution requirements

Principal provides investment services through Genesys Cloud CX, a cloud-based contact center that provides powerful, native integrations with AWS. Each year, Principal handles millions of calls and digital interactions. As a first step, they wanted to transcribe voice calls and analyze those interactions to determine primary call drivers, including issues, topics, sentiment, average handle time (AHT) breakdowns, and develop additional natural language processing (NLP)-based analytics.

In order analyze the calls properly, Principal had a few requirements:

• Contact details: Understanding the customer journey requires understanding whether a speaker is an automated interactive voice response (IVR) system or a human agent and when a call transfer occurs between the two.
• Content redaction: Each customer audio interaction is recorded as a stereo WAV file, but could potentially include sensitive information such as HIPAA-protected and personally identifiable information (PII).
• Scalability: This architecture needed to immediately scale to thousands of calls per day and millions of calls per year. In addition, Principal needed an extensible analytics architecture that analyze other channels such as email threads and traditional voice of the customer (VoC) survey results.
• Integrity is non-negotiable at Principal—it guides everything they do. In fact, doing what’s right is one of the core values at Principal. Therefore, when the Principal team started tackling this project, they knew that ensuring the highest standard of data security such as regulatory compliance, data privacy, and data quality would be a non-negotiable, key requirement. The team needed to utilize technology with a matching stance on data security, and the ability to build custom compliance and security controls to uphold strict requirements. Attention to this key requirement allows Principal to maintain a safe and secure customer experience.

Solution overview

After extensive research, the Principal team finalized AWS Contact Center Intelligence (CCI) solutions, which empower companies to improve customer experience and gain conversation insights by adding AI capabilities to third-party on-premises and cloud contact centers. The CCI Post-Call Analytics (PCA) solution is part of CCI solutions suite and fit many of the identified requirements. PCA has a Solutions Library Guidance reference architecture with an open-source example repository on GitHub. Working with their AWS account team, Principal detailed the PCA solution and its deployment, and set up custom training programs and immersion days to rapidly upskill the Principal teams. The example architecture (see the following diagram) and code base in the open-source repository allowed the Principal engineering teams to jumpstart their solution around unifying the customer journey, and merging telephony records and transcript records together.

PCA provides an entire architecture around ingesting audio files in a fully automated workflow with AWS Step Functions, which is initiated when an audio file is delivered to a configured Amazon Simple Storage Service (Amazon S3) bucket. After a few minutes, a transcript is produced with Amazon Transcribe Call Analytics and saved to another S3 bucket for processing by other business intelligence (BI) tools. PCA also offers a web-based user interface that allows customers to browse call transcripts. PCA’s security features ensure that any PII data was redacted from the transcript, as well as from the audio file itself. Additionally, all data within the S3 bucket can be encrypted with keys belonging to Principal.

Principal worked with AWS technical teams to modify the Step Functions workflow within PCA to further achieve their goals. Call details such as interaction timestamps, call queues, agent transfers, and participant speaking times are tracked by Genesys in a file called a Contact Trace Record (CTR). Combining accurate transcripts with Genesys CTR files, Principal could properly identify the speakers, categorize the calls into groups, analyze agent performance, identify upsell opportunities, and conduct additional machine learning (ML)-powered analytics.

The teams built a new data ingestion mechanism, allowing the CTR files to be jointly delivered with the audio file to an S3 bucket. Principal and AWS collaborated on a new AWS Lambda function that was added to the Step Functions workflow. This Lambda function identifies CTR records and provides an additional processing step that outputs an enhanced transcript containing additional metadata such as queue and agent ID information, IVR identification and tagging, and how many agents (and IVRs) the customer was transferred to, all aggregated from the CTR records. This extra information enables Principal to create a map of the customer interaction throughout the lifecycle of the conversation and focus on the critical speech segments, while excluding less relevant ones.

Additionally, this postprocessing step enabled Principal to further enrich transcripts with internal information such as agent and queue names and expand the analytics capabilities of PCA, including custom NLP-based ML models for topic and customer intent identification, deployed using Amazon SageMaker endpoints, and additional transcript augmentation using foundational generative AI models hosted on Amazon Bedrock.

PCA is open source on GitHub, which allows customers such as Principal to extend and maintain their own forks with customized, private business code. It also allows the community to submit code back to the main repository for others to use. Principal and AWS technical teams partnered to merge the Genesys CTR and postprocessing placeholder features into the main release of PCA. This partnership between Principal and AWS enabled speed-to-market for Principal, while ensuring that existing and incoming business requirements could be rapidly added. The contributions to the open-source project has accelerated other customers’ Genesys CTR workloads.

Answer business questions

Once PCA was in place, Principal analysts, data scientists, engineers, and business owners worked with AWS SMEs to build numerous Amazon QuickSight dashboards to display the data insights and begin answering business questions. QuickSight is a cloud-scale BI service that you can use to deliver easy-to-understand insights from multiple datasets, from AWS data, third-party data, software as a service (SaaS) data, and more. The use of this BI tool, with its native integrations to the existing data repositories made accessible by Amazon Athena, made the creation of visualizations to display the large-scale data relatively straightforward, and enabled self-service BI. Visualizations were quickly drafted to answer some key questions, including “What are our customers calling us about,” “What topics relate to the longest AHT/most transfers,” and “What topics and issues relate to the lowest customer sentiment scores?” By ingesting additional data related to Principal custom topic models, the team was able to expand their use of QuickSight to include topic and correlation comparisons, model validation capabilities, and comparisons of sentiment based on speaker, segment, call, and conversation. In addition, the use of QuickSight insights quickly allowed the Principal teams to implement anomaly detection and volume prediction, while Amazon QuickSight Q, an ML feature within QuickSight that uses NLP, enabled rapid natural language quantitative data analytics.

When the initial initiative for PCA was complete, Principal knew they needed to immediately dive deeper into the omnichannel customer experience. Together, Principal and AWS have built data ingestion pipelines for customer email interactions and additional metadata from their customer data platform, and built data aggregation and analytics mechanisms to combine omnichannel data into a single customer insight lens. Utilization of Athena views and QuickSight dashboards has continued to enable classic analytics, and the implementation of proof of concept graph databases via Amazon Neptune will help Principal extract insights into interaction topics and intent relationships within the omnichannel view when implemented at scale.

The Results

PCA helped accelerate time to market. Principal was able to deploy the existing open-source PCA app by themselves in 1 day. Then, Principal worked together with AWS and expanded the PCA offering with numerous features like the Genesys CTR integration over a period of 3 months. The development and deployment process was a joint, iterative process that allowed Principal to test and process production call volumes on newly built features. Since the initial engagement, AWS and Principal continue to work together, sharing business requirements, roadmaps, code, and bug fixes to expand PCA.

Since its initial development and deployment, Principal has processed over 1 million customer calls through the PCA framework. This resulted in over 63 million individual speech segments spoken by a customer, agent, or IVR. With this wealth of data, Principal has been able to conduct large-scale historical and near-real-time analytics to gain insights into the customer experience.

AWS CCI solutions are a game-changer for Principal. Principal’s existing suite of CCI tools, which includes Qualtrics for simple dashboarding and opportunity identification, was expanded with the addition of PCA. The addition of PCA to the suite of CCI tools enabled Principal to rapidly conduct deep analytics on their contact center interactions. With this data, Principal now can conduct advanced analytics to understand customer interactions and call drivers, including topics, intents, issues, action items, and outcomes. Even in a small-scale, controlled production environment, the PCA data lake has spawned numerous new use cases.

Roadmap

The data generated from PCA could be used to make critical business decisions regarding call routing based on insights around which topics are driving longer average handle time, longer holds, more transfers, and negative customer sentiment. Knowledge on when customer interactions with the IVR and automated voice assistants are misunderstood or misrouted will help Principal improve the self-service experience. Understanding why a customer called instead of using the website is critical to improving the customer journey and boosting customer happiness. Product managers responsible for enhancing web experiences have shared how excited they are to be able to use data from PCA to drive their prioritization of new enhancements and measure the impact of changes. Principal is also analyzing other potential use cases such as customer profile mapping, fraud detection, workforce management, the use of additional AI/ML and large language models (LLMs), and identifying new and emerging trends within their contact centers.

In the future, Principal plans to continue expanding postprocessing capabilities with additional data aggregation, analytics, and natural language generation (NLG) models for text summarization. Principal is currently integrating generative AI and foundational models (such as Amazon Titan) to their proprietary solutions. Principal plans to use AWS generative AI to enhance employee productivity, grow assets under management, deliver high-quality customer experiences, and deliver tools that allow customers to make investment and retirement decisions efficiently. Given the flexibility and extensibility of the open-source PCA framework, the teams at Principal have an extensive list of additional enhancements, analytics, and insights that could extend the existing framework.

“With AWS Post Call analytics solution, Principal can currently conduct large-scale historical analytics to understand where customer experiences can be improved, generate actionable insights, and prioritize where to act. Now, we are adding generative AI using Amazon Bedrock to help our business users make data-driven decisions with higher speed and accuracy, while reducing costs. We look forward to exploring the post call summarization feature in Amazon Transcribe Call Analytics in order to enable our agents to focus their time and resources engaging with customers, rather than manual after contact work.”

– says Miguel Sanchez Urresty, Director of Data & Analytics at Principal Financial Group.

Conclusion

The AWS CCI PCA solution is designed to improve customer experience, derive customer insights, and reduce operational costs by adding AI and ML to the contact center provider of your choice. To learn more about other CCI solutions, such as Live Call Analytics, refer to AWS Contact Center Intelligence (CCI) Solutions.

About Principal Financial Group

Principal Financial Group and affiliates, Des Moines IA is a financial company with 19,000 employees. In business for more than 140 years, we’re helping more than 62 million customers in various countries around the world as of December 31, 2022.

AWS and Amazon are not affiliates of any company of the Principal Financial Group Insurance products issued by Principal National Life Insurance Co (except in NY) and Principal Life Insurance Company. Plan administrative services offered by Principal Life. Principal Funds, Inc. is distributed by Principal Funds Distributor, Inc. Securities offered through Principal Securities, Inc., member SIPC and/or independent broker/dealers. Referenced companies are members of the Principal Financial Group, Des Moines, IA 50392. ©2023 Principal Financial Services, Inc.

This communication is intended to be educational in nature and is not intended to be taken as a recommendation. Insurance products and plan administrative services provided through Principal Life Insurance Company, a member of the Principal Financial Group, Des Moines, IA 50392

About the authors

Christopher Lott is a Senior Solutions Architect in the AWS AI Language Services team. He has 20 years of enterprise software development experience. Chris lives in Sacramento, California, and enjoys gardening, cooking, aerospace/general aviation, and traveling the world.

Dr. Nicki Susman is a Senior Data Scientist and the Technical Lead of the Principal Language AI Services team. She has extensive experience in data and analytics, application development, infrastructure engineering, and DevSecOps.

Read More

Prompt engineering has become an essential skill for anyone working with large language models (LLMs) to generate high-quality and relevant texts. Although text prompt engineering has been widely discussed, visual prompt engineering is an emerging field that requires attention. Visual prompts can include bounding boxes or masks that guide vision models in generating relevant and accurate outputs. In this post, we explore the basics of visual prompt engineering, its benefits, and how it can be used to solve a specific use case: image segmentation for autonomous driving.

In recent years, the field of computer vision has witnessed significant advancements in the area of image segmentation. One such breakthrough is the Segment Anything Model (SAM) by Meta AI, which has the potential to revolutionize object-level segmentation with zero-shot or few-shot training. In this post, we use the SAM model as an example foundation vision model and explore its application to the BDD100K dataset, a diverse autonomous driving dataset for heterogeneous multitask learning. By combining the strengths of SAM with the rich data provided by BDD100K, we showcase the potential of visual prompt engineering with different versions of SAM. Inspired by the LangChain framework for language models, we propose a visual chain to perform visual prompting by combining object detection models with SAM.

Although this post focuses on autonomous driving, the concepts discussed are applicable broadly to domains that have rich vision-based applications such as healthcare and life sciences, and media and entertainment. Let’s begin by learning a little more about what’s under the hood of a foundational vision model like SAM. We used Amazon SageMaker Studio on an ml.g5.16xlarge instance for this post.

Segment Anything Model (SAM)

Foundation models are large machine learning (ML) models trained on vast quantity of data and can be prompted or fine-tuned for task-specific use cases. Here, we explore the Segment Anything Model (SAM), which is a foundational model for vision, specifically image segmentation. It is pre-trained on a massive dataset of 11 million images and 1.1 billion masks, making it the largest segmentation dataset as of writing. This extensive dataset covers a wide range of objects and categories, providing SAM with a diverse and large-scale training data source.

The SAM model is trained to understand objects and can output segmentation masks for any object in images or video frames. The model allows for visual prompt engineering, enabling you to provide inputs such as text, points, bounding boxes, or masks to generate labels without altering the original image. SAM is available in three sizes: base (ViT-B, 91 million parameters), large (ViT-L, 308 million parameters), and huge (ViT-H, 636 million parameters), catering to different computational requirements and use cases.

The primary motivation behind SAM is to improve object-level segmentation with minimal training samples and epochs for any objects of interest. The power of SAM lies in its ability to adapt to new image distributions and tasks without prior knowledge, a feature known as zero-shot transfer. This adaptability is achieved through its training on the expansive SA-1B dataset, which has demonstrated impressive zero-shot performance, surpassing many prior fully supervised results.

As shown in the following architecture for SAM, the process of generating segmentation masks involves three steps:

1. An image encoder produces a one-time embedding for the image.
2. A prompt encoder converts any prompt into an embedding vector for the prompt.
3. The lightweight decoder combines the information from the image encoder and the prompt encoder to predict segmentation masks.

As an example, we can provide an input with an image and bounding box around an object of interest in that image (e.g. Silver car or driving lane)  and SAM model would produce segmentation masks for that object.

Visual prompt engineering

Prompt engineering refers to structuring inputs to a model that makes the model understand the intent and produces desired outcome. With textual prompt engineering, you can structure the input text through modifications such as choice of words, formatting, ordering, and more to get the desired output. Visual prompt engineering assumes that the user is working in a visual modality (image or video), and provides inputs. The following is a non-exhaustive list of potential ways to provide input to the generative AI model in the visual domain:

• Point – A singular (x, y) coordinate point in the image plane
• Points – Multiple (x, y) coordinate points, not necessarily related to each other
• Bounding box – A set of four values (x, y, w, h) that define a rectangular region in the image plane
• Contour – A set of (x, y) coordinate points in the image plane that form a closed shape
• Mask – An array the same size as the image with a partial mask of the object of interest

With the visual prompt engineering techniques in mind, let’s explore how this can be applied to the SAM pre-trained model. We have use the base version of the pre-trained model.

Zero-shot prompting with the pre-trained SAM model

To start with, let’s explore the zero-shot approach. The following is a sample image from the training dataset taken from a vehicle’s front camera.

We can get segmentation masks for all objects from the image without any explicit visual prompting by automatically generating masks with just an input image. In the following image, we see parts of the car, road, traffic sign, license plates, flyover, pillars, signs, and more are segmented.

However, this output is not immediately useful for the following reasons:

• The cars are not segmented as a whole, but in parts. For most perception models, for example, we don’t really care about each of the tires having separate output masks. This is true when looking for other known objects of interest as well, such as road, vegetation, signs, and so on.
• Parts of the image that are useful for downstream tasks like drivable area are split up, with no explanation. On the other hand, similar instances are identified separately, and we may be interested in grouping similar objects (panoptic vs. instance segmentation).

Visual prompt engineering with the pre-trained SAM model

Fortunately, SAM supports providing input prompts, and we can use points, point arrays, and bounding boxes as inputs. With these specific instructions, we expect SAM to do better with segmentations focusing on specific points or areas. This can be compared with the language prompt template
"What is a good name for a company that makes {product}?"
where the input along with this prompt template from the user is the {product}. {product} is an input slot. In visual prompting, the bounding boxes, points, or masks are the input slots.

The following image provides the original ground truth bounding box around vehicles, and the drivable area patch from BDD100K ground truth data. The image also shows an input point (a yellow X) at the center of the green bounding box that we will refer to in the next few sections.

Let’s try to generate a mask for the car on the left with the green bounding box as an input to SAM. As shown in the following example, the base model of SAM doesn’t really find anything. This is also seen in the low segmentation score. When we look at the segmentation masks more closely, we see that there are small regions returned as masks (pointed at using red arrows) that aren’t really usable for any downstream application.

Let’s try a combination of a bounding box and a point as the input visual prompt. The yellow cross in the preceding image is the center of the bounding box. Providing this point’s (x,y) coordinates as the prompt along with the bounding box constraint gives us the following mask and a slightly higher score. This is still not usable by any means.

Finally, with the base pre-trained model, we can provide just the input point as a prompt (without the bounding box). The following images show two of the top three masks we thought were interesting.

Mask 1 segments the full car, whereas Mask 3 segments out an area that holds the car’s number plate close to the yellow cross (input prompt). Mask 1 is still not a tight, clean mask around the car; this points to the quality of the model, which we can assume increases with model size.

We can try larger pre-trained models with the same input prompt. The following images show our results. When using the huge SAM pre-trained model, Mask 3 is the entire car, whereas Mask 1 and 2 can be used to extract the number plate.

The large version of the SAM model also provides similar outputs.

The process we went through here is similar to manual prompt engineering for text prompts that you may already be familiar with. Note that a recent improvement in the SAM model to segment anything in high quality provides much better object- and context-specific outputs. In our case, we find that zero-shot prompting with text and visual prompts (point, box, and point and box inputs) don’t improve results drastically as we saw above.

Prompt templates and visual chains

As we can see from the preceding zero-shot examples, SAM struggles to identify all the objects in the scene. This is a good example of where we can take advantage of prompt templates and visual chains. Visual chain is inspired by the chain concept in the popular LangChain framework for language applications. It helps chain the data sources and an LLM to produce the output. For example, we can use an API chain to call an API and invoke an LLM to answer the question based on the API response.

Inspired by LangChain, we propose a sequential visual chain that looks like the following figure. We use a tool (like a pre-trained object detection model) to get initial bounding boxes, calculate the point at the center of the bounding box, and use this to prompt the SAM model with the input image.

For example, the following image shows the segmentation masks as a result of running this chain.

Another example chain can involve a text input of the object the user is interested in identifying. To implement this, we built a pipeline using Grounding DINO, an object detection model to prompt SAM for segmentation.

Grounding DINO is a zero-shot object detection model that can perform object detection with text providing category names (such as “traffic lights” or “truck”) and expressions (such as “yellow truck”). It accepts pairs of text and image to perform the object detection. It’s based on a transformer architecture and enables cross modalities with text and image data. To learn more about Grounding DINO, refer to Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. This generates bounding boxes and labels and can be processed further to generate center points, filter based on labels, thresholds, and more. This is used (boxes or points) as a prompt to SAM for segmentation, which outputs masks.

The following are some examples showing the input text, DINO output (bounding boxes), and the final SAM output (segmentation masks).

The following images show the output for “yellow truck.”

The following images show the output for “silver car.”

The following image shows the output for “driving lane.”

We can use this pipeline to build a visual chain. The following code snippet explains this concept:

pipeline = [object_predictor, segment_predictor]
image_chain = ImageChain.from_visual_pipeline(pipeline, image_store, verbose=True)
image_chain.run('All silver cars', image_id='5X3349')

Although this is a simple example, this concept can be extended to process feeds from cameras on vehicles to perform object tracking, personally identifiable information (PII) data redaction, and more. We can also get the bounding boxes from smaller models, or in some cases, using standard computer vision tools. It’s fairly straightforward to use a pre-trained model or a service like Amazon Rekognition to get initial (visual) labels for your prompt. At the time of writing this, there are over 70 models available on Amazon SageMaker Jumpstart for object detection, and Amazon Rekognition already identifies several useful categories of objects in images, including cars, pedestrians, and other vehicles.

Next, we look at some quantitative results related to performance of SAM models with a subset of BDD100K data.

Quantitative results

Our objective is to compare the performance of three pre-trained models when given the same visual prompting. In this case, we use the center point of the object location as the visual input. We compare the performance with respect to the object sizes (in proportion to image size)— small (area <0.11%), medium (0.11% < area < 1%), and large (area > 1%). The bounding box area thresholds are defined by the Common Objects in Context (COCO) evaluation metrics [Lin et al., 2014].

The evaluation is at the pixel level and we use the following evaluation metrics:

• Precision = (number relevant and retrieved instances) / (total number of retrieved instances)
• Recall = (number of relevant and retrieve instances) / (total number of relevant instances)
• Instances here are each pixel within the bounding box of the object of interest

The following table reports the performance of three different versions of the SAM model (base, large, and huge). These versions have three different encoders: ViT-B (base), ViT-L (large), ViT-H (huge). The encoders have different parameter counts, where the base model has less parameters than large, and large is less than huge. Although increasing the number of parameters shows improved performance with larger objects, this is not so for smaller objects.

Fine-tuning SAM for your use case

In many cases, directly using a pre-trained SAM model may not be very useful. For example, let’s look at a typical scene in traffic—the following picture is the output from the SAM model with randomly sampled prompt points as input on the left, and the actual labels from the semantic segmentation task from BDD100K on the right. These are obviously very different.

Perception stacks in AVs can easily use the second image, but not the first. On the other hand, there some useful outputs from the first image that can be used, and that the model was not explicitly trained on, for example, lane markings, sidewalk segmentation, license plate masks, and so on. We can fine-tune the SAM model to improve the segmentation results. To perform this fine-tuning, we created a training dataset using an instance segmentation subset (500 images) from the BDD10K dataset. This is a very small subset of images, but our purpose is to prove that foundational vision models (much like LLMs) can perform well for your use case with a surprisingly small number of images. The following image shows the input image, output mask (in blue, with a red border for the car on the left), and possible prompts (bounding box in green and center point X in yellow).

We performed fine-tuning using the Hugging Face library on Amazon SageMaker Studio. We used the ml.g4dn.xlarge instance for the SAM base model tests, and the ml.g4dn.2xlarge for the SAM huge model tests. In our initial experiments, we observed that fine-tuning the base model with just bounding boxes was not successful. The fine-tuned and pre-trained models weren’t able to learn car-specific ground truth masks from the original datasets. Adding query points to the fine-tuning also didn’t improve the training.

Next, we can try fine-tuning the SAM huge model for 30 epochs, with a very small dataset (500 images). The original ground truth mask looks like the following image for the label type car.

As shown in the following images, the original pre-trained version of the huge model with a specific bounding box prompt (in green) gives no output, whereas the fine-tuned version gives an output (still not accurate but fine-tuning was cut off after 40 epochs, and with a very small training dataset of 500 images). The original, pre-trained huge model wasn’t able to predict masks for any of the images we tested. As an example downstream application, the fine-tuned model can be used in pre-labeling workflows such as the one described in Auto-labeling module for deep learning-based Advanced Driver Assistance Systems on AWS.

Conclusion

In this post, we discussed the foundational vision model known as the Segment Anything Model (SAM) and its architecture. We used the SAM model to discuss visual prompting and the various inputs to visual prompting engineering. We explored how different visual prompts perform and their limitations. We also described how visual chains increase performance over using just one prompt, similar to the LangChain API. Next, we provided a quantitative evaluation of three pre-trained models. Lastly, we discussed the fine-tuned SAM model and its results compared to the original base model. Fine-tuning of foundation models helps improve model performance for specific tasks like segmentation. It should be noted that SAM model due to its resource requirements, limits usage for real-time use-cases and inferencing at the edge in its current state. We hope with future iterations and improved techniques, would reduce compute requirements and improve latency.

It is our hope that this post encourages you to explore visual prompting for your use cases. Because this is still an emerging form of prompt engineering, there is much to discover in terms of visual prompts, visual chains, and performance of these tools. Amazon SageMaker is a fully managed ML platform that enables builders to explore large language and visual models and build generative AI applications. Start building the future with AWS today.

About the authors

Gopi Krishnamurthy is a Senior AI/ML Solutions Architect at Amazon Web Services based in New York City. He works with large Automotive customers as their trusted advisor to transform their Machine Learning workloads and migrate to the cloud. His core interests include deep learning and serverless technologies. Outside of work, he likes to spend time with his family and explore a wide range of music.

Shreyas Subramanian is a Principal AI/ML specialist Solutions Architect, and helps customers by using Machine Learning to solve their business challenges using the AWS platform. Shreyas has a background in large scale optimization and Machine Learning, and in use of Machine Learning and Reinforcement Learning for accelerating optimization tasks.

 Sujitha Martin is an Applied Scientist in the Generative AI Innovation Center (GAIIC). Her expertise is in building machine learning solutions involving computer vision and natural language processing for various industry verticals. In particular, she has extensive experience working on human-centered situational awareness and knowledge infused learning for highly autonomous systems.

Francisco Calderon is a Data Scientist in the Generative AI Innovation Center (GAIIC). As a member of the GAIIC, he helps discover the art of the possible with AWS customers using Generative AI technologies. In his spare time, Francisco likes to play music and guitar, playing soccer with his daughters, and enjoying time with his family.

Read More

Online communities are driving user engagement across industries like gaming, social media, ecommerce, dating, and e-learning. Members of these online communities trust platform owners to provide a safe and inclusive environment where they can freely consume content and contribute. Content moderators are often employed to review user-generated content and check that it’s safe and compliant with your terms of use. However, the ever-increasing scale, complexity, and variety of inappropriate content makes human moderation workflows unscalable and expensive. The result is poor, harmful, and non-inclusive communities that disengage users and negatively impact the community and business.

Along with user-generated content, machine-generated content has brought a fresh challenge to content moderation. It automatically creates highly realistic content that may be inappropriate or harmful at scale. The industry is facing the new challenge of automatically moderating content generated by AI to protect users from harmful material.

In this post, we introduce toxicity detection, a new feature from Amazon Comprehend that helps you automatically detect harmful content in user- or machine-generated text. This includes plain text, text extracted from images, and text transcribed from audio or video content.

Detect toxicity in text content with Amazon Comprehend

Amazon Comprehend is a natural-language processing (NLP) service that uses machine learning (ML) to uncover valuable insights and connections in text. It offers a range of ML models that can be either pre-trained or customized through API interfaces. Amazon Comprehend now provides a straightforward, NLP-based solution for toxic content detection in text.

The Amazon Comprehend Toxicity Detection API assigns an overall toxicity score to text content, ranging from 0–1, indicating the likelihood of it being toxic. It also categorizes text into the following seven categories and provides a confidence score for each:

• HATE_SPEECH – Speech that criticizes, insults, denounces, or dehumanizes a person or a group on the basis of an identity, be it race, ethnicity, gender identity, religion, sexual orientation, ability, national origin, or another identity group.
• GRAPHIC – Speech that uses visually descriptive, detailed, and unpleasantly vivid imagery. Such language is often made verbose so as to amplify an insult, or discomfort or harm to the recipient.
• HARASSMENT_OR_ABUSE – Speech that imposes disruptive power dynamics between the speaker and hearer (regardless of intent), seeks to affect the psychological well-being of the recipient, or objectifies a person.
• SEXUAL – Speech that indicates sexual interest, activity, or arousal by using direct or indirect references to body parts, physical traits, or sex.
• VIOLENCE_OR_THREAT – Speech that includes threats that seek to inflict pain, injury, or hostility towards a person or group.
• INSULT – Speech that includes demeaning, humiliating, mocking, insulting, or belittling language.
• PROFANITY – Speech that contains words, phrases, or acronyms that are impolite, vulgar, or offensive.

You can access the Toxicity Detection API by calling it directly using the AWS Command Line Interface (AWS CLI) and AWS SDKs. Toxicity detection in Amazon Comprehend is currently supported in the English language.

Use cases

Text moderation plays a crucial role in managing user-generated content across diverse formats, including social media posts, online chat messages, forum discussions, website comments, and more. Moreover, platforms that accept video and audio content can use this feature to moderate transcribed audio content.

The emergence of generative AI and large language models (LLMs) represents the latest trend in the field of AI. Consequently, there is a growing need for responsive solutions to moderate content generated by LLMs. The Amazon Comprehend Toxicity Detection API is ideally suited for addressing this need.

Amazon Comprehend Toxicity Detection API request

You can send up to 10 text segments to the Toxicity Detection API, each with a size limit of 1 KB. Every text segment in the request is handled independently. In the following example, we generate a JSON file named toxicity_api_input.json containing the text content, including three sample text segments for moderation. Note that in the example, the profane words are masked as XXXX.

{
  "TextSegments": [     
    {"Text": "and go through the door go through the door he's on the right"},
    {"Text": "he's on the right XXXXX him"},
    {"Text": "what the XXXX are you doing man that's why i didn't want to play"}
  ],
  "LanguageCode": "en"
}

You can use the AWS CLI to invoke the Toxicity Detection API using the preceding JSON file containing the text content:

aws comprehend detect-toxic-content --cli-input-json file://toxicity_api_input.json

Amazon Comprehend Toxicity Detection API response

The Toxicity Detection API response JSON output will include the toxicity analysis result in the ResultList field. ResultList lists the text segment items, and the sequence represents the order in which the text sequences were received in the API request. Toxicity represents the overall confidence score of detection (between 0–1). Labels includes a list of toxicity labels with confidence scores, categorized by the type of toxicity.

The following code shows the JSON response from the Toxicity Detection API based on the request example in the previous section:

{
    "ResultList": [
        {
            "Toxicity": 0.009200000204145908,
            "Labels": [
                { "Name": "PROFANITY", "Score": 0.0007999999797903001},
                { "Name": "HATE_SPEECH", "Score": 0.0017999999690800905},
                { "Name": "INSULT", "Score": 0.003000000026077032},
                { "Name": "GRAPHIC", "Score": 0.0010000000474974513},
                { "Name": "HARASSMENT_OR_ABUSE", "Score": 0.0013000000035390258},
                { "Name": "SEXUAL", "Score": 0.0017000000225380063},
                { "Name": "VIOLENCE_OR_THREAT", "Score": 0.004999999888241291}
            ]
        },
        {
            "Toxicity": 0.7358999848365784,
            "Labels": [
                { "Name": "PROFANITY", "Score": 0.011900000274181366},
                { "Name": "HATE_SPEECH", "Score": 0.019500000402331352},
                { "Name": "INSULT", "Score": 0.0714000016450882},
                { "Name": "GRAPHIC", "Score": 0.006099999882280827},
                { "Name": "HARASSMENT_OR_ABUSE", "Score": 0.018200000748038292},
                { "Name": "SEXUAL", "Score": 0.0027000000700354576},
                { "Name": "VIOLENCE_OR_THREAT", "Score": 0.8145999908447266}
            ]
        },
        {
            "Toxicity": 0.9843000173568726,
            "Labels": [
                { "Name": "PROFANITY", "Score": 0.9369999766349792 },
                { "Name": "HATE_SPEECH", "Score": 0.30880001187324524 },
                { "Name": "INSULT", "Score": 0.42100000381469727 },
                { "Name": "GRAPHIC", "Score": 0.12630000710487366 },
                { "Name": "HARASSMENT_OR_ABUSE", "Score": 0.25519999861717224 },
                { "Name": "SEXUAL", "Score": 0.19169999659061432 },
                { "Name": "VIOLENCE_OR_THREAT", "Score": 0.19539999961853027 }
            ]
        }
    ]
}

In the preceding JSON, the first text segment is considered safe with a low toxicity score. However, the second and third text segments received toxicity scores of 73% and 98%, respectively. For the second segment, Amazon Comprehend detects a high toxicity score for VIOLENCE_OR_THREAT; for the third segment, it detects PROFANITY with a high toxicity score.

Sample request using the Python SDK

The following code snippet demonstrates how to utilize the Python SDK to invoke the Toxicity Detection API. This code receives the same JSON response as the AWS CLI command demonstrated earlier.

import boto3 import base64
# Initialize a Comprehend boto3 client object
comprehend_client = session.client('comprehend')

# Call comprehend Detect Toxic Content API with text segments
response = comprehend_client.detect_toxic_content(
    TextSegments=[
        {"Text":  "and go through the door go through the door he's on the right"},
        {"Text":  "he's on the right XXXXX him"},
        {"Text":  "what the XXXX are you doing man that's why i didn't want to play"}
    ],
    LanguageCode='en'
)

Summary

In this post, we provided an overview of the new Amazon Comprehend Toxicity Detection API. We also described how you can parse the API response JSON. For more information, refer to Comprehend API document.

Amazon Comprehend toxicity detection is now generally available in four Regions: us-east-1, us-west-2, eu-west-1, and ap-southeast-2.

To learn more about content moderation, refer to Guidance for Content Moderation on AWS. Take the first step towards streamlining your content moderation operations with AWS.

About the Authors

Lana Zhang is a Senior Solutions Architect at AWS WWSO AI Services team, specializing in AI and ML for Content Moderation, Computer Vision, Natural Language Processing and Generative AI. With her expertise, she is dedicated to promoting AWS AI/ML solutions and assisting customers in transforming their business solutions across diverse industries, including social media, gaming, e-commerce, media, advertising & marketing.

Ravisha SK is a Senior Product Manager, Technical at AWS with a focus on AI/ML. She has over 10 years of experience in data analytics and machine learning across different domains. In her spare time, she enjoys reading, experimenting in the kitchen and exploring new coffee shops.

Read More

Today, we are excited to announce the capability to fine-tune the Mistral 7B model using Amazon SageMaker JumpStart. You can now fine-tune and deploy Mistral text generation models on SageMaker JumpStart using the Amazon SageMaker Studio UI with a few clicks or using the SageMaker Python SDK.

Foundation models perform very well with generative tasks, from crafting text and summaries, answering questions, to producing images and videos. Despite the great generalization capabilities of these models, there are often use cases that have very specific domain data (such as healthcare or financial services), and these models may not be able to provide good results for these use cases. This results in a need for further fine-tuning of these generative AI models over the use case-specific and domain-specific data.

In this post, we demonstrate how to fine-tune the Mistral 7B model using SageMaker JumpStart.

What is Mistral 7B

Mistral 7B is a foundation model developed by Mistral AI, supporting English text and code generation abilities. It supports a variety of use cases, such as text summarization, classification, text completion, and code completion. To demonstrate the customizability of the model, Mistral AI has also released a Mistral 7B-Instruct model for chat use cases, fine-tuned using a variety of publicly available conversation datasets.

Mistral 7B is a transformer model and uses grouped query attention and sliding window attention to achieve faster inference (low latency) and handle longer sequences. Grouped query attention is an architecture that combines multi-query and multi-head attention to achieve output quality close to multi-head attention and comparable speed to multi-query attention. The sliding window attention method uses the multiple levels of a transformer model to focus on information that came earlier, which helps the model understand a longer stretch of context. . Mistral 7B has an 8,000-token context length, demonstrates low latency and high throughput, and has strong performance when compared to larger model alternatives, providing low memory requirements at a 7B model size. The model is made available under the permissive Apache 2.0 license, for use without restrictions.

You can fine-tune the models using either the SageMaker Studio UI or SageMaker Python SDK. We discuss both methods in this post.

Fine-tune via the SageMaker Studio UI

In SageMaker Studio, you can access the Mistral model via SageMaker JumpStart under Models, notebooks, and solutions, as shown in the following screenshot.

If you don’t see Mistral models, update your SageMaker Studio version by shutting down and restarting. For more information about version updates, refer to Shut down and Update Studio Apps.

On the model page, you can point to the Amazon Simple Storage Service (Amazon S3) bucket containing the training and validation datasets for fine-tuning. In addition, you can configure deployment configuration, hyperparameters, and security settings for fine-tuning. You can then choose Train to start the training job on a SageMaker ML instance.

Deploy the model

After the model is fine-tuned, you can deploy it using the model page on SageMaker JumpStart. The option to deploy the fine-tuned model will appear when fine-tuning is complete, as shown in the following screenshot.

Fine-tune via the SageMaker Python SDK

You can also fine-tune Mistral models using the SageMaker Python SDK. The complete notebook is available on GitHub. In this section, we provide examples of two types of fine-tuning.

Instruction fine-tuning

Instruction tuning is a technique that involves fine-tuning a language model on a collection of natural language processing (NLP) tasks using instructions. In this technique, the model is trained to perform tasks by following textual instructions instead of specific datasets for each task. The model is fine-tuned with a set of input and output examples for each task, allowing the model to generalize to new tasks that it hasn’t been explicitly trained on as long as prompts are provided for the tasks. Instruction tuning helps improve the accuracy and effectiveness of models and is helpful in situations where large datasets aren’t available for specific tasks.

Let’s walk through the fine-tuning code provided in the example notebook with the SageMaker Python SDK.

We use a subset of the Dolly dataset in an instruction tuning format, and specify the template.json file describing the input and the output formats. The training data must be formatted in JSON lines (.jsonl) format, where each line is a dictionary representing a single data sample. In this case, we name it train.jsonl.

The following snippet is an example of train.jsonl. The keys instruction, context, and response in each sample should have corresponding entries {instruction}, {context}, {response} in the template.json.

{
    "instruction": "What is a dispersive prism?", 
    "context": "In optics, a dispersive prism is an optical prism that is used to disperse light, that is, to separate light into its spectral components (the colors of the rainbow). Different wavelengths (colors) of light will be deflected by the prism at different angles. This is a result of the prism material's index of refraction varying with wavelength (dispersion). Generally, longer wavelengths (red) undergo a smaller deviation than shorter wavelengths (blue). The dispersion of white light into colors by a prism led Sir Isaac Newton to conclude that white light consisted of a mixture of different colors.", 
    "response": "A dispersive prism is an optical prism that disperses the light's different wavelengths at different angles. When white light is shined through a dispersive prism it will separate into the different colors of the rainbow."
}

The following is a sample of template.json:

{
    "prompt": "Below is an instruction that describes a task, paired with an input that provides further context. "
    "Write a response that appropriately completes the request.nn"
    "### Instruction:n{instruction}nn### Input:n{context}nn",
    "completion": " {response}",
}

After you upload the prompt template and the training data to an S3 bucket, you can set the hyperparameters.

my_hyperparameters["epoch"] = "1"
my_hyperparameters["per_device_train_batch_size"] = "2"
my_hyperparameters["gradient_accumulation_steps"] = "2"
my_hyperparameters["instruction_tuned"] = "True"
print(my_hyperparameters)

You can then start the fine-tuning process and deploy the model to an inference endpoint. In the following code, we use an ml.g5.12xlarge instance:

from sagemaker.jumpstart.estimator import JumpStartEstimator

instruction_tuned_estimator = JumpStartEstimator(
    model_id=model_id,
    hyperparameters=my_hyperparameters,
    instance_type="ml.g5.12xlarge",
)
instruction_tuned_estimator.fit({"train": train_data_location}, logs=True)

instruction_tuned_predictor = instruction_tuned_estimator.deploy()

Domain adaptation fine-tuning

Domain adaptation fine-tuning is a process that refines a pre-trained LLM to better suit a specific domain or task. By using a smaller, domain-specific dataset, the LLM can be fine-tuned to understand and generate content that is more accurate, relevant, and insightful for that specific domain, while still retaining the vast knowledge it gained during its original training.

The Mistral model can be fine-tuned on any domain-specific dataset. After it’s fine-tuned, it’s expected to generate domain-specific text and solve various NLP tasks in that specific domain. For the training dataset, provide a train directory and an optional validation directory, each containing a single CSV, JSON, or TXT file. For CSV and JSON formats, use data from the text column or the first column if text isn’t present. Ensure only one file exists under each directory. For instance, input data may be SEC filings of Amazon as a text file:

This report includes estimates, projections, statements relating to our
business plans, objectives, and expected operating results that are “forward-
looking statements” within the meaning of the Private Securities Litigation
Reform Act of 1995, Section 27A of the Securities Act of 1933, and Section 21E
of the Securities Exchange Act of 1934. Forward-looking statements may appear
throughout this report, including the following sections: “Business” (Part I,
Item 1 of this Form 10-K), “Risk Factors” (Part I, Item 1A of this Form 10-K),
and “Management’s Discussion and Analysis of Financial Condition and Results
of Operations” (Part II, Item 7 of this Form 10-K). These forward-looking
statements generally are identified by the words “believe,” “project,”
“expect,” “anticipate,” “estimate,” “intend,” “strategy,” “future,”
“opportunity,” “plan,” “may,” “should,” “will,” “would,” “will be,” “will
continue,” “will likely result,” and similar expressions.

You can start domain adaptation fine-tuning by specifying the hyperparameter “instruction_tuned” as “False“. The rest of the steps are similar to the instruction fine-tuning steps.

We set max_input_length to be 2048 on g5.12xlarge. You can use a higher input length on a larger instance type. For details, see the example notebook.

Inference example with and without fine-tuning

The following table contains the results of the Mistral 7B model fine-tuned with SEC filing documents of Amazon from 2021–2022. We compare the output before and after fine-tuning.

As you can see, the fine-tuned model provides more specific information related to Amazon compared to the generic pre-trained one. This is because fine-tuning adapts the model to understand the nuances, patterns, and specifics of the provided dataset. By using a pre-trained model and tailoring it with fine-tuning, we ensure that you get the best of both worlds: the broad knowledge of the pre-trained model and the specialized accuracy for your unique dataset. One size may not fit all in the world of machine learning, and fine-tuning is the tailor-made solution you need!

Conclusion

In this post, we discussed fine-tuning the Mistral 7B model using SageMaker JumpStart. We showed how you can use the SageMaker JumpStart console in SageMaker Studio or the SageMaker Python SDK to fine-tune and deploy these models. As a next step, you can try fine-tuning these models on your own dataset using the code provided in the GitHub repository to test and benchmark the results for your use cases.

About the Authors

Xin Huang is a Senior Applied Scientist for Amazon SageMaker JumpStart and Amazon SageMaker built-in algorithms. He focuses on developing scalable machine learning algorithms. His research interests are in the area of natural language processing, explainable deep learning on tabular data, and robust analysis of non-parametric space-time clustering. He has published many papers in ACL, ICDM, KDD conferences, and Royal Statistical Society: Series A.

Vivek Gangasani is a AI/ML Startup Solutions Architect for Generative AI startups at AWS. He helps emerging GenAI startups build innovative solutions using AWS services and accelerated compute. Currently, he is focused on developing strategies for fine-tuning and optimizing the inference performance of Large Language Models. In his free time, Vivek enjoys hiking, watching movies and trying different cuisines.

Dr. Ashish Khetan is a Senior Applied Scientist with Amazon SageMaker built-in algorithms and helps develop machine learning algorithms. He got his PhD from University of Illinois Urbana-Champaign. He is an active researcher in machine learning and statistical inference, and has published many papers in NeurIPS, ICML, ICLR, JMLR, ACL, and EMNLP conferences.

Read More

Fake news, defined as news that conveys or incorporates false, fabricated, or deliberately misleading information, has been around as early as the emergence of the printing press. The rapid spread of fake news and disinformation online is not only deceiving to the public, but can also have a profound impact on society, politics, economy, and culture. Examples include:

• Cultivating distrust in the media
• Undermining the democratic process
• Spreading false or discredited science (for example, the anti-vax movement)

Advances in artificial intelligence (AI) and machine learning (ML) have made developing tools for creating and sharing fake news even easier. Early examples include advanced social bots and automated accounts that supercharge the initial stage of spreading fake news. In general, it is not trivial for the public to determine whether such accounts are people or bots. In addition, social bots are not illegal tools, and many companies legally purchase them as part of their marketing strategy. Therefore, it’s not easy to curb the use of social bots systematically.

Recent discoveries in the field of generative AI make it possible to produce textual content at an unprecedented pace with the help of large language models (LLMs). LLMs are generative AI text models with over 1 billion parameters, and they are facilitated in the synthesis of high-quality text.

In this post, we explore how you can use LLMs to tackle the prevalent issue of detecting fake news. We suggest that LLMs are sufficiently advanced for this task, especially if improved prompt techniques such as Chain-of-Thought and ReAct are used in conjunction with tools for information retrieval.

We illustrate this by creating a LangChain application that, given a piece of news, informs the user whether the article is true or fake using natural language. The solution also uses Amazon Bedrock, a fully managed service that makes foundation models (FMs) from Amazon and third-party model providers accessible through the AWS Management Console and APIs.

LLMs and fake news

The fake news phenomenon started evolving rapidly with the advent of the internet and more specifically social media (Nielsen et al., 2017). On social media,­ fake news can be shared quickly in a user’s network, leading the public to form the wrong collective opinion. In addition, people often propagate fake news impulsively, ignoring the factuality of the content if the news resonates with their personal norms (Tsipursky et al. 2018). Research in social science has suggested that cognitive bias (confirmation bias, bandwagon effect, and choice-supportive bias) is one of the most pivotal factors in making irrational decisions in terms of the both creation and consumption of fake news (Kim, et al., 2021). This also implies that news consumers share and consume information only in the direction of strengthening their beliefs.

The power of generative AI to produce textual and rich content at an unprecedented pace aggravates the fake news problem. An example worth mentioning is deepfake technology—combining various images on an original video and generating a different video. Besides the disinformation intent that human actors bring to the mix, LLMs add a whole new set of challenges:

• Factual errors – LLMs have an increased risk of containing factual errors due to the nature of their training and ability to be creative while generating the next words in a sentence. LLM training is based on repeatedly presenting a model with incomplete input, then using ML training techniques until it correctly fills in the gaps, thereby learning language structure and a language-based world model. Consequently, although LLMs are great pattern matchers and re-combiners (“stochastic parrots”), they fail at a number of simple tasks that require logical reasoning or mathematical deduction, and can hallucinate answers. In addition, temperature is one of the LLM input parameters that controls the behavior of the model when generating the next word in a sentence. By selecting a higher temperature, the model will use a lower-probability word, providing a more random response.
• Lengthy – Generated texts tend to be lengthy and lack a clearly defined granularity for individual facts.
• Lack of fact-checking – There is no standardized tooling available for fact-checking during the process of text generation.

Overall, the combination of human psychology and limitations of AI systems has created a perfect storm for the proliferation of fake news and misinformation online.

Solution overview

LLMs are demonstrating outstanding capabilities in language generation, understanding, and few-shot learning. They are trained on a vast corpus of text from the internet, where quality and accuracy of extracted natural language may not be assured.

In this post, we provide a solution to detect fake news based both on Chain-of-Thought and Re-Act (Reasoning and Acting) prompt approaches. First, we discuss those two prompt engineering techniques, then we show their implementation using LangChain and Amazon Bedrock.

The following architecture diagram outlines the solution for our fake news detector.

Architecture diagram for fake news detection.

We use a subset of the FEVER dataset containing a statement and the ground truth about the statement indicating false, true, or unverifiable claims (Thorne J. et al., 2018).

The workflow can be broken down into the following steps:

1. The user selects one of the statements to check if fake or true.
2. The statement and the fake news detection task are incorporated into the prompt.
3. The prompt is passed to LangChain, which invokes the FM in Amazon Bedrock.
4. Amazon Bedrock returns a response to the user request with the statement True or False.

In this post, we use the Claude v2 model from Anthrophic (anthropic.claude-v2). Claude is a generative LLM based on Anthropic’s research into creating reliable, interpretable, and steerable AI systems. Created using techniques like constitutional AI and harmlessness training, Claude excels at thoughtful dialogue, content creation, complex reasoning, creativity, and coding. However, by using Amazon Bedrock and our solution architecture, we also have the flexibility to choose among other FMs provided by Amazon, AI21labs, Cohere, and Stability.ai.

You can find the implementation details in the following sections. The source code is available in the GitHub repository.

Prerequisites

For this tutorial, you need a bash terminal with Python 3.9 or higher installed on either Linux, Mac, or a Windows Subsystem for Linux and an AWS account.

We also recommend using either an Amazon SageMaker Studio notebook, an AWS Cloud9 instance, or an Amazon Elastic Compute Cloud (Amazon EC2) instance.

Deploy fake news detection using the Amazon Bedrock API

The solution uses the Amazon Bedrock API, which can be accessed using the AWS Command Line Interface (AWS CLI), the AWS SDK for Python (Boto3), or an Amazon SageMaker notebook. Refer to the Amazon Bedrock User Guide for more information. For this post, we use the Amazon Bedrock API via the AWS SDK for Python.

Set up Amazon Bedrock API environment

To set up your Amazon Bedrock API environment, complete the following steps:

1. Download the latest Boto3 or upgrade it:

   pip install --upgrade boto3

2. Make sure you configure the AWS credentials using the aws configure command or pass them to the Boto3 client.
3. Install the latest version of LangChain:

   pip install “langchain>=0.0.317” --quiet

You can now test your setup using the following Python shell script. The script instantiates the Amazon Bedrock client using Boto3. Next, we call the list_foundation_models API to get the list of foundation models available for use.

import boto3 
import json 
bedrock = boto3.client( 'bedrock', region_name=YOUR_REGION) 
print(json.dumps(bedrock.list_foundation_models(), indent=4))

After successfully running the preceding command, you should get the list of FMs from Amazon Bedrock.

LangChain as a prompt chaining solution

To detect fake news for a given sentence, we follow the zero-shot Chain-of-Thought reasoning process (Wei J. et al., 2022), which is composed of the following steps:

1. Initially, the model attempts to create a statement about the news prompted.
2. The model creates a bullet point list of assertions.
3. For each assertion, the model determines if the assertion is true or false. Note that using this methodology, the model relies exclusively on its internal knowledge (weights computed in the pre-training phase) to reach a verdict. The information is not verified against any external data at this point.
4. Given the facts, the model answers TRUE or FALSE for the given statement in the prompt.

To achieve these steps, we use LangChain, a framework for developing applications powered by language models. This framework allows us to augment the FMs by chaining together various components to create advanced use cases. In this solution, we use the built-in SimpleSequentialChain in LangChain to create a simple sequential chain. This is very useful, because we can take the output from one chain and use it as the input to another.

Amazon Bedrock is integrated with LangChain, so you only need to instantiate it by passing the model_id when instantiating the Amazon Bedrock object. If needed, the model inference parameters can be provided through the model_kwargs argument, such as:

• maxTokenCount – The maximum number of tokens in the generated response
• stopSequences – The stop sequence used by the model
• temperature – A value that ranges between 0–1, with 0 being the most deterministic and 1 being the most creative
• top – A value that ranges between 0–1, and is used to control tokens’ choices based on the probability of the potential choices

If this is the first time you are using an Amazon Bedrock foundational model, make sure you request access to the model by selecting from the list of models on the Model access page on the Amazon Bedrock console, which in our case is claude-v2 from Anthropic.

from langchain.llms.bedrock import Bedrock
bedrock_runtime = boto3.client(
    service_name='bedrock-runtime',
    region_name= YOUR_REGION,
)
model_kwargs={
        'max_tokens_to_sample': 8192
    }
llm = Bedrock(model_id=" anthropic.claude-v2", client=bedrock_runtime, model_kwargs=model_kwargs)

The following function defines the Chain-of-Thought prompt chain we mentioned earlier for detecting fake news. The function takes the Amazon Bedrock object (llm) and the user prompt (q) as arguments. LangChain’s PromptTemplate functionality is used here to predefine a recipe for generating prompts.

from langchain.prompts import PromptTemplate
from langchain.chains import LLMChain
from langchain.chains import SimpleSequentialChain

def generate_and_print(llm, q):
    total_prompt = """"""

    # the model is asked to create a bullet point list of assertions
    template = """Here is a statement:
    {statement}
    Make a bullet point list of the assumptions you made when given the above statement.nn"""
    prompt_template = PromptTemplate(input_variables=["statement"], template=template)
    assumptions_chain = LLMChain(llm=llm, prompt=prompt_template)
    total_prompt = total_prompt + template

    # the model is asked to create a bullet point list of assertions    
    template = """Here is a bullet point list of assertions:
    {assertions}
    For each assertion, determine whether it is true or false. If it is false, explain why.nn"""
    prompt_template = PromptTemplate(input_variables=["assertions"], template=template)
    fact_checker_chain = LLMChain(llm=llm, prompt=prompt_template)
    total_prompt = total_prompt + template

    #for each assertion, the model is askded to determine if the assertion is true or false, based on internal knowledge alone

    template = """ Based on the above assertions, the final response is FALSE if one of the assertions is FALSE. Otherwise, the final response is TRUE. You should only respond with TRUE or FALSE.'{}'""".format(q)
    template = """{facts}n""" + template
    prompt_template = PromptTemplate(input_variables=["facts"], template=template)
    answer_chain = LLMChain(llm=llm, prompt=prompt_template)
    total_prompt = total_prompt + template

    #SimpleSequentialChain allows us to take the output from one chain and use it as the input to another
    overall_chain = SimpleSequentialChain(chains=[assumptions_chain, fact_checker_chain, answer_chain], verbose=True)
    answer = overall_chain.run(q)

    return answer

The following code calls the function we defined earlier and provides the answer. The statement is TRUE or FALSE. TRUE means that the statement provided contains correct facts, and FALSE means that the statement contains at least one incorrect fact.

from IPython.display import display, Markdown

q="The first woman to receive a Ph.D. in computer science was Dr. Barbara Liskov, who earned her degree from Stanford University in 1968."
print(f'The statement is: {q}')
display(Markdown(generate_and_print(llm, q)))

An example of a statement and model response is provided in the following output:

The statement is: The first woman to receive a Ph.D. in computer science was Dr. Barbara Liskov, who earned her degree from Stanford University in 1968.

> Entering new SimpleSequentialChain chain...
 Here is a bullet point list of assumptions I made about the statement:

- Dr. Barbara Liskov was the first woman to earn a Ph.D. in computer science. 

- Dr. Liskov earned her Ph.D. from Stanford University.

- She earned her Ph.D. in 1968.

- No other woman earned a Ph.D. in computer science prior to 1968.

- Stanford University had a computer science Ph.D. program in 1968. 

- The statement refers to Ph.D. degrees earned in the United States.
 Here are my assessments of each assertion:

- Dr. Barbara Liskov was the first woman to earn a Ph.D. in computer science.
  - True. Dr. Liskov was the first American woman to earn a Ph.D. in computer science, which she received from Stanford University in 1968.

- Dr. Liskov earned her Ph.D. from Stanford University.
  - True. Multiple sources confirm she received her Ph.D. from Stanford in 1968.

- She earned her Ph.D. in 1968.
  - True. This is consistent across sources.

- No other woman earned a Ph.D. in computer science prior to 1968.
  - False. While she was the first American woman, Mary Kenneth Keller earned a Ph.D. in computer science from the University of Wisconsin in 1965. However, Keller earned her degree in the US as well.

- Stanford University had a computer science Ph.D. program in 1968.
  - True. Stanford established its computer science department and Ph.D. program in 1965.

- The statement refers to Ph.D. degrees earned in the United States.
  - False. The original statement does not specify the country. My assumptions that it refers to the United States is incorrect. Keller earned her Ph.D. in the US before Liskov.
 False

ReAct and tools

In the preceding example, the model correctly identified that the statement is false. However, submitting the query again demonstrates the model’s inability to distinguish the correctness of facts. The model doesn’t have the tools to verify the truthfulness of statements beyond its own training memory, so subsequent runs of the same prompt can lead it to mislabel fake statements as true. In the following code, you have a different run of the same example:

The statement is: The first woman to receive a Ph.D. in computer science was Dr. Barbara Liskov, who earned her degree from Stanford University in 1968.

> Entering new SimpleSequentialChain chain...
 Here is a bullet point list of assumptions I made about the statement:

- Dr. Barbara Liskov was the first woman to earn a Ph.D. in computer science
- Dr. Liskov earned her Ph.D. degree in 1968 
- Dr. Liskov earned her Ph.D. from Stanford University
- Stanford University awarded Ph.D. degrees in computer science in 1968
- Dr. Liskov was a woman
- Ph.D. degrees existed in 1968
- Computer science existed as a field of study in 1968
 Here are my assessments of each assertion:

- Dr. Barbara Liskov was the first woman to earn a Ph.D. in computer science
    - True. Dr. Liskov was the first woman to earn a Ph.D. in computer science in 1968 from Stanford University.

- Dr. Liskov earned her Ph.D. degree in 1968
    - True. Multiple sources confirm she received her Ph.D. in computer science from Stanford in 1968.

- Dr. Liskov earned her Ph.D. from Stanford University 
    - True. Dr. Liskov earned her Ph.D. in computer science from Stanford University in 1968.

- Stanford University awarded Ph.D. degrees in computer science in 1968
    - True. Stanford awarded Liskov a Ph.D. in computer science in 1968, so they offered the degree at that time.

- Dr. Liskov was a woman
    - True. All biographical information indicates Dr. Liskov is female.

- Ph.D. degrees existed in 1968
    - True. Ph.D. degrees have existed since the late 19th century.

- Computer science existed as a field of study in 1968
    - True. While computer science was a relatively new field in the 1960s, Stanford and other universities offered it as a field of study and research by 1968.
 True

One technique for guaranteeing truthfulness is ReAct. ReAct (Yao S. et al., 2023) is a prompt technique that augments the foundation model with an agent’s action space. In this post, as well as in the ReAct paper, the action space implements information retrieval using search, lookup, and finish actions from a simple Wikipedia web API.

The reason behind using ReAct in comparison to Chain-of-Thought is to use external knowledge retrieval to augment the foundation model to detect if a given piece of news is fake or true.

In this post, we use LangChain’s implementation of ReAct through the agent ZERO_SHOT_REACT_DESCRIPTION. We modify the previous function to implement ReAct and use Wikipedia by using the load_tools function from the langchain.agents.

We also need to install the Wikipedia package:

Below is the new code:

from langchain.agents import load_tools, initialize_agent, AgentType

def generate_and_print(llm, q):
    print(f'Inside generate_and_print: q = {q}')
    tools = load_tools(["wikipedia"], llm=llm)
    agent = initialize_agent(tools, llm, 
                             agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION, 
                             verbose=True,
                             handle_parsing_errors=True,
                             agent_kwargs={})

    input = """Here is a statement:
    {statement}
    Is this statement correct? You can use tools to find information if needed.
    The final response is FALSE if the statement is FALSE. Otherwise, TRUE."""

    answer = agent.run(input.format(statement=q))
    
    return answer

The following is the output of the preceding function given the same statement used before:

> Entering new AgentExecutor chain...
 Here are my thoughts and actions to determine if the statement is true or false:

Thought: To verify if this statement about the first woman to receive a PhD in computer science is true, I should consult a reliable information source like Wikipedia.

Action: Wikipedia
Action Input: first woman to receive phd in computer science
Observation: Page: Fu Foundation School of Engineering and Applied Science
Summary: The Fu Foundation School of Engineering and Applied Science (popularly known as SEAS or Columbia Engineering; previously known as Columbia School of Mines) is the engineering and applied science school of Columbia University. It was founded as the School of Mines in 1863 and then the School of Mines, Engineering and Chemistry before becoming the School of Engineering and Applied Science. On October 1, 1997, the school was renamed in honor of Chinese businessman Z.Y. Fu, who had donated $26 million to the school.
The Fu Foundation School of Engineering and Applied Science maintains a close research tie with other institutions including NASA, IBM, MIT, and The Earth Institute. Patents owned by the school generate over $100 million annually for the university. SEAS faculty and alumni are responsible for technological achievements including the developments of FM radio and the maser.
The School's applied mathematics, biomedical engineering, computer science and the financial engineering program in operations research are very famous and highly ranked. The current SEAS faculty include 27 members of the National Academy of Engineering and one Nobel laureate. In all, the faculty and alumni of Columbia Engineering have won 10 Nobel Prizes in physics, chemistry, medicine, and economics.
The school consists of approximately 300 undergraduates in each graduating class and maintains close links with its undergraduate liberal arts sister school Columbia College which shares housing with SEAS students. The School's current dean is Shih-Fu Chang, who was appointed in 2022.

Page: Doctor of Science
Summary: A Doctor of Science (Latin: Scientiae Doctor; most commonly abbreviated DSc or ScD) is an academic research doctorate awarded in a number of countries throughout the world. In some countries, a Doctor of Science is the degree used for the standard doctorate in the sciences; elsewhere a Doctor of Science is a "higher doctorate" awarded in recognition of a substantial and sustained contribution to scientific knowledge beyond that required for a Doctor of Philosophy (PhD).

Page: Timeline of women in science
Summary: This is a timeline of women in science, spanning from ancient history up to the 21st century. While the timeline primarily focuses on women involved with natural sciences such as astronomy, biology, chemistry and physics, it also includes women from the social sciences (e.g. sociology, psychology) and the formal sciences (e.g. mathematics, computer science), as well as notable science educators and medical scientists. The chronological events listed in the timeline relate to both scientific achievements and gender equality within the sciences.
Thought: Based on the Wikipedia pages, the statement appears to be false. The Wikipedia Timeline of Women in Science page indicates that Adele Goldstine was the first woman to earn a PhD in computer science in 1964 from the University of Michigan, not Barbara Liskov from Stanford in 1968. Therefore, my final answer is:

Final Answer: FALSE

Clean up

To save costs, delete all the resources you deployed as part of the tutorial. If you launched AWS Cloud9 or an EC2 instance, you can delete it via the console or using the AWS CLI. Similarly, you can delete the SageMaker notebook you may have created via the SageMaker console.

Limitations and related work

The field of fake news detection is actively researched in the scientific community. In this post, we used Chain-of-Thought and ReAct techniques and in evaluating the techniques, we only focused on the accuracy of the prompt technique classification (if a given statement is true or false). Therefore, we haven’t considered other important aspects such as speed of the response, nor extended the solution to additional knowledge base sources besides Wikipedia.

Although this post focused on two techniques, Chain-of-Thought and ReAct, an extensive body of work has explored how LLMs can detect, eliminate or mitigate fake news. Lee et al. has proposed the use of an encoder-decoder model using NER (named entity recognition) to mask the named entities in order to ensure that the token masked actually uses the knowledge encoded in the language model. Chern et.al. developed FacTool, which uses Chain-of-Thought principles to extract claims from the prompt, and consequently collect relevant evidences of the claims. The LLM then judges the factuality of the claim given the retrieved list of evidences. Du E. et al. presents a complementary approach where multiple LLMs propose and debate their individual responses and reasoning processes over multiple rounds in order to arrive at a common final answer.

Based on the literature, we see that the effectiveness of LLMs in detecting fake news increases when the LLMs are augmented with external knowledge and multi-agent conversation capability. However, these approaches are more computationally complex because they require multiple model calls and interactions, longer prompts, and lengthy network layer calls. Ultimately, this complexity translates into an increased overall cost. We recommend assessing the cost-to-performance ratio before deploying similar solutions in production.

Conclusion

In this post, we delved into how to use LLMs to tackle the prevalent issue of fake news, which is one of the major challenges of our society nowadays. We started by outlining the challenges presented by fake news, with an emphasis on its potential to sway public sentiment and cause societal disruptions.

We then introduced the concept of LLMs as advanced AI models that are trained on a substantial quantity of data. Due to this extensive training, these models boast an impressive understanding of language, enabling them to produce human-like text. With this capacity, we demonstrated how LLMs can be harnessed in the battle against fake news by using two different prompt techniques, Chain-of-Thought and ReAct.

We underlined how LLMs can facilitate fact-checking services on an unparalleled scale, given their capability to process and analyze vast amounts of text swiftly. This potential for real-time analysis can lead to early detection and containment of fake news. We illustrated this by creating a Python script that, given a statement, highlights to the user whether the article is true or fake using natural language.

We concluded by underlining the limitations of the current approach and ended on a hopeful note, stressing that, with the correct safeguards and continuous enhancements, LLMs could become indispensable tools in the fight against fake news.

We’d love to hear from you. Let us know what you think in the comments section, or use the issues forum in the GitHub repository.

Disclaimer: The code provided in this post is meant for educational and experimentation purposes only. It should not be relied upon to detect fake news or misinformation in real-world production systems. No guarantees are made about the accuracy or completeness of fake news detection using this code. Users should exercise caution and perform due diligence before utilizing these techniques in sensitive applications.

To get started with Amazon Bedrock, visit the Amazon Bedrock console.

About the authors

Anamaria Todor is a Principal Solutions Architect based in Copenhagen, Denmark. She saw her first computer when she was 4 years old and never let go of computer science, video games, and engineering since. She has worked in various technical roles, from freelancer, full-stack developer, to data engineer, technical lead, and CTO, at various companies in Denmark, focusing on the gaming and advertising industries. She has been at AWS for over 3 years, working as a Principal Solutions Architect, focusing mainly on life sciences and AI/ML. Anamaria has a bachelor’s in Applied Engineering and Computer Science, a master’s degree in Computer Science, and over 10 years of AWS experience. When she’s not working or playing video games, she’s coaching girls and female professionals in understanding and finding their path through technology.

Marcel Castro is a Senior Solutions Architect based in Oslo, Norway. In his role, Marcel helps customers with architecture, design, and development of cloud-optimized infrastructure. He is a member of the AWS Generative AI Ambassador team with the goal to drive and support EMEA customers on their generative AI journey. He holds a PhD in Computer Science from Sweden and a master’s and bachelor’s degree in Electrical Engineering and Telecommunications from Brazil.

Read More

In the era of big data and AI, companies are continually seeking ways to use these technologies to gain a competitive edge. One of the hottest areas in AI right now is generative AI, and for good reason. Generative AI offers powerful solutions that push the boundaries of what’s possible in terms of creativity and innovation. At the core of these cutting-edge solutions lies a foundation model (FM), a highly advanced machine learning model that is pre-trained on vast amounts of data. Many of these foundation models have shown remarkable capability in understanding and generating human-like text, making them a valuable tool for a variety of applications, from content creation to customer support automation.

However, these models are not without their challenges. They are exceptionally large and require large amounts of data and computational resources to train. Additionally, optimizing the training process and calibrating the parameters can be a complex and iterative process, requiring expertise and careful experimentation. These can be barriers for many organizations looking to build their own foundation models. To overcome this challenge, many customers are considering to fine-tune existing foundation models. This is a popular technique to adjust a small portion of model parameters for specific applications while still preserving the knowledge already encoded in the model. It allows organizations to use the power of these models while reducing the resources required to customize to a specific domain or task.

There are two primary approaches to fine-tuning foundation models: traditional fine-tuning and parameter-efficient fine-tuning. Traditional fine-tuning involves updating all the parameters of the pre-trained model for a specific downstream task. On the other hand, parameter-efficient fine-tuning includes a variety of techniques that allow for customization of a model without updating all the original model parameters. One such technique is called Low-rank Adaptation (LoRA). It involves adding small, task-specific modules to the pre-trained model and training them while keeping the rest of the parameters fixed as shown in the following image.

Source: Generative AI on AWS (O’Reilly, 2023)

LoRA has gained popularity recently for several reasons. It offers faster training, reduced memory requirements, and the ability to reuse pre-trained models for multiple downstream tasks. More importantly, the base model and adapter can be stored separately and combined at any time, making it easier to store, distribute, and share fine-tuned versions. However, this introduces a new challenge: how to properly manage these new types of fine-tuned models. Should you combine the base model and adapter or keep them separate? In this post, we walk through best practices for managing LoRA fine-tuned models on Amazon SageMaker to address this emerging question.

Working with FMs on SageMaker Model Registry

In this post, we walk through an end-to-end example of fine-tuning the Llama2 large language model (LLM) using the QLoRA method. QLoRA combines the benefits of parameter efficient fine-tuning with 4-bit/8-bit quantization to further reduce the resources required to fine-tune a FM to a specific task or use case. For this, we will use the pre-trained 7 billion parameter Llama2 model and fine-tune it on the databricks-dolly-15k dataset. LLMs like Llama2 have billions of parameters and are pretrained on massive text datasets. Fine-tuning adapts an LLM to a downstream task using a smaller dataset. However, fine-tuning large models is computationally expensive. This is why we will use the QLoRA method to quantize the weights during finetuning to reduce this computation cost.

In our examples, you will find two notebooks (llm-finetune-combined-with-registry.ipynb and llm-finetune-separate-with-registry.ipynb). Each works through a different way to handle LoRA fine-tuned models as illustrated in the following diagram:

1. First, we download the pre-trained Llama2 model with 7 billion parameters using SageMaker Studio Notebooks. LLMs, like Llama2, have shown state-of-the-art performance on natural language processing (NLP) tasks when fine-tuned on domain-specific data.
2. Next, we fine-tune Llama2 on the databricks-dolly-15k dataset using the QLoRA method. QLoRA reduces the computational cost of fine-tuning by quantizing model weights.
3. During fine-tuning, we integrate SageMaker Experiments Plus with the Transformers API to automatically log metrics like gradient, loss, etc.
4. We then version the fine-tuned Llama2 model in SageMaker Model Registry using two approaches:
   1. Storing the full model
   2. Storing the adapter and base model separately.
5. Finally, we host the fine-tuned Llama2 models using Deep Java Library (DJL) Serving on a SageMaker Real-time endpoint.

In the following sections, we will dive deeper into each of these steps, to demonstrate the flexibility of SageMaker for different LLM workflows and how these features can help improve the operations of your models.

Prerequisites

Complete the following prerequisites to start experimenting with the code.

• Create a SageMaker Studio Domain: Amazon SageMaker Studio, specifically Studio Notebooks, is used to kick off the Llama2 fine-tuning task then register and view models within SageMaker Model Registry. SageMaker Experiments is also used to view and compare Llama2 fine-tuning job logs (training loss/test loss/etc.).
• Create an Amazon Simple Storage Service (S3) bucket: Access to an S3 bucket to store training artifacts and model weights is required. For instructions, refer to Creating a bucket. The sample code used for this post will use the SageMaker default S3 bucket but you can customize it to use any relevant S3 bucket.
• Set up Model Collections (IAM permissions): Update your SageMaker Execution Role with permissions to resource-groups as listed under Model Registry Collections Developer Guide to implement Model Registry grouping using Model Collections.
• Accept the Terms & Conditions for Llama2: You will need to accept the end-user license agreement and acceptable use policy for using the Llama2 foundation model.

The examples are available in the GitHub repository. The notebook files are tested using Studio notebooks running on PyTorch 2.0.0 Python 3.10 GPU Optimized kernel and ml.g4dn.xlarge instance type.

Experiments plus callback integration

Amazon SageMaker Experiments lets you organize, track, compare and evaluate machine learning (ML) experiments and model versions from any integrated development environment (IDE), including local Jupyter Notebooks, using the SageMaker Python SDK or boto3. It provides the flexibility to log your model metrics, parameters, files, artifacts, plot charts from the different metrics, capture various metadata, search through them and support model reproducibility. Data scientists can quickly compare the performance and hyperparameters for model evaluation through visual charts and tables. They can also use SageMaker Experiments to download the created charts and share the model evaluation with their stakeholders.

Training LLMs can be a slow, expensive, and iterative process. It is very important for a user to track LLM experimentation at scale to prevent an inconsistent model tuning experience. HuggingFace Transformer APIs allow users to track metrics during training tasks through Callbacks. Callbacks are “read only” pieces of code that can customize the behavior of the training loop in the PyTorch Trainer that can inspect the training loop state for progress reporting, logging on TensorBoard or SageMaker Experiments Plus via custom logic (which is included as a part of this codebase).

You can import the SageMaker Experiments callback code included in this post’s code repository as shown in the following code block:

# imports a custom implementation of Experiments Callback
from smexperiments_callback import SageMakerExperimentsCallback
...
...
# Create Trainer instance with SageMaker experiments callback
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=validation_dataset,
    data_collator=default_data_collator,
    callbacks=[SageMakerExperimentsCallback] # Add our Experiments Plus Callback function
)

This callback will automatically log the following information into SageMaker Experiments as a part of the training run:

• Training Parameters and Hyper-Parameters
• Model Training and Validation loss at Step, Epoch and Final
• Model Input and Output artifacts (training dataset, validation dataset, model output location, training debugger and more)

The following graph shows examples of the charts you can display by using that information.

This allows you to compare multiple runs easily using the Analyze feature of SageMaker Experiments. You can select the experiment runs you want to compare, and they will automatically populate comparison graphs.

Register fine-tuned models to Model Registry Collections

Model Registry Collections is a feature of SageMaker Model Registry that allows you to group registered models that are related to each other and organize them in hierarchies to improve model discoverability at scale. We will use Model Registry Collections to keep track of the base model and fine-tuned variants.

Full Model Copy method

The first method combines the base model and LoRA adapter and saves the full fine-tuned model. The following code illustrates the model merging process and saves the combined model using model.save_pretrained().

if args.merge_weights:
        
    trainer.model.save_pretrained(temp_dir, safe_serialization=False)
    # clear memory
    del model
    del trainer
    torch.cuda.empty_cache()
    
    from peft import AutoPeftModelForCausalLM

    # load PEFT model in fp16
    model = AutoPeftModelForCausalLM.from_pretrained(
        temp_dir,
        low_cpu_mem_usage=True,
        torch_dtype=torch.float16,
    )  
    # Merge LoRA and base model and save
    model = model.merge_and_unload()        
    model.save_pretrained(
        args.sm_model_dir, safe_serialization=True, max_shard_size="2GB"
    )

Combining the LoRA adapter and base model into a single model artifact after fine-tuning has advantages and disadvantages. The combined model is self-contained and can be independently managed and deployed without needing the original base model. The model can be tracked as its own entity with a version name reflecting the base model and fine-tuning data. We can adopt a nomenclature using the base_model_name + fine-tuned dataset_name to organize the model groups. Optionally, model collections could associate the original and fine-tuned models, but this may not be necessary since the combined model is independent.  The following code snippet shows you how to register the fine-tuned model.

# Model Package Group Vars
ft_package_group_name = f"{model_id.replace('/', '--')}-{dataset_name}"
ft_package_group_desc = "QLoRA for model Mikael110/llama-2-7b-{dataset_name}-fp16"
...
...
...
model_package_group_input_dict = {
    "ModelPackageGroupName" : ft_package_group_name,
    "ModelPackageGroupDescription" : ft_package_group_desc,
    "Tags": ft_tags
}
create_model_pacakge_group_response = sm_client.create_model_package_group(
**model_package_group_input_dict
)

You can use the training estimator to register the model into Model Registry.

inference_image_uri = sagemaker.image_uris.retrieve(
    "djl-deepspeed", region=region, version="0.23.0"
)
print(f"Image going to be used is ---- > {inference_image_uri}")

model_package = huggingface_estimator.register(
    content_types=["application/json"],
    response_types=["application/json"],
    inference_instances=[
        "ml.p2.16xlarge", 
...
...
...
    ],
    image_uri = inference_image_uri,
    customer_metadata_properties = {"training-image-uri": huggingface_estimator.training_image_uri()},  #Store the training image url
    model_package_group_name=ft_model_pkg_group_name,
    approval_status="Approved"
)

model_package_arn = model_package.model_package_arn
print("Model Package ARN : ", model_package_arn)

From Model Registry, you can retrieve the model package and deploy that model directly.

endpoint_name = f"{name_from_base(model_group_for_base)}-endpoint"

model_package.deploy(
    initial_instance_count=1,
    instance_type="ml.g5.12xlarge",
    endpoint_name=endpoint_name
)

However, there are drawbacks to this approach. Combining the models leads to storage inefficiency and redundancy since the base model is duplicated in each fine-tuned version. As model size and the number of fine-tuned models increase, this exponentially inflates storage needs. Taking the llama2 7b model as an example, the base model is approximately 13 GB and the fine-tuned model is 13.6 GB. 96% percent of the model needs to be duplicated after each fine tuning. Additionally, distributing and sharing very large model files also becomes more difficult and presents operational challenges as file transfer and management cost increases with increasing model size and fine-tune jobs.

Separate adapter and base method

The second method focuses on separation of base weights and adapter weights by saving them as separate model components and loading them sequentially at runtime.

    ..
    ..
    ..
    else:   
        # save finetuned LoRA model and then the tokenizer for inference
        trainer.model.save_pretrained(
            args.sm_model_dir, 
            safe_serialization=True
        )
    tokenizer.save_pretrained(
        args.sm_model_dir
    )

Saving base and adapter weights has advantages and disadvantages, similar to the Full Model Copy method. One advantage is that it can save storage space. The base weights, which are the largest component of a fine-tuned model, are only saved once and can be reused with other adapter weights that are tuned for different tasks. For example, the base weights of Llama2-7B are about 13 GB, but each fine-tuning task only needs to store about 0.6 GB of adapter weights, which is a 95% space savings. Another advantage is that base weights can be managed separately from adapter weights using a base weights only model registry. This can be useful for SageMaker domains that are running in a VPC only mode without an internet gateway, since the base weights can be accessed without having to go through the internet.

Create Model Package Group for base weights

### Create Model Package Group
base_package_group_name = model_id.replace('/', '--')
base_package_group_desc = "Source: https://huggingface.co/Mikael110/llama-2-7b-guanaco-fp16"
...
...
...
model_package_group_input_dict = {
    "ModelPackageGroupName" : base_package_group_name,
    "ModelPackageGroupDescription" : base_package_group_desc,
    "Tags": base_tags
}
create_model_pacakge_group_response = sm_client.create_model_package_group(
**model_package_group_input_dict
)

>>>
Created ModelPackageGroup Arn : arn:aws:sagemaker:us-west-2:376678947624:model-package-group/Mikael110--llama-2-7b-guanaco-fp16
...
...
...

### Register Base Model Weights
from sagemaker.huggingface import HuggingFaceModel

# create Hugging Face Model Class
huggingface_model = HuggingFaceModel(
    transformers_version='4.28',
    pytorch_version='2.0',
    py_version='py310',
    model_data=model_data_uri, # this is an S3 path to your base weights as *.tar.gz
    role=role,
)

_response = huggingface_model.register(
    content_types=["application/json"],
    response_types=["application/json"],
    inference_instances=[
    "ml.p2.16xlarge",
    ...
    ],
    transform_instances=[
    "ml.p2.16xlarge",
    ...
    ],
    model_package_group_name=base_model_pkg_group_name,
    approval_status="Approved"
 )

Create Model Package Group for QLoRA weights

The following code shows how to tag QLoRA weights with the dataset/task type and register fine-tuned delta weights into a separate model registry and track the delta weights separately.

### Create Model Package Group for delta weights
ft_package_group_name = f"{model_id.replace('/', '--')}-finetuned-sql"
ft_package_group_desc = "QLoRA for model Mikael110/llama-2-7b-guanaco-fp16"
ft_tags = [
    {
    "Key": "modelType",
    "Value": "QLoRAModel"
    },
    {
    "Key": "fineTuned",
    "Value": "True"
    },
    {
    "Key": "sourceDataset",
    "Value": f"{dataset_name}"
    }
]
model_package_group_input_dict = {
    "ModelPackageGroupName" : ft_package_group_name,
    "ModelPackageGroupDescription" : ft_package_group_desc,
    "Tags": ft_tags
}
create_model_pacakge_group_response = sm_client.create_model_package_group(
**model_package_group_input_dict
)
print(f'Created ModelPackageGroup Arn : {create_model_pacakge_group_response["ModelPackageGroupArn"]}')
ft_model_pkg_group_name = create_model_pacakge_group_response["ModelPackageGroupArn"]

>>> 
Created ModelPackageGroup Arn : arn:aws:sagemaker:us-east-1:811828458885:model-package-group/mikael110--llama-2-7b-guanaco-fp16-finetuned-sql

...
...
...

### Register Delta Weights QLoRA Model Weights
huggingface_model = HuggingFaceModel(
    transformers_version='4.28',
    pytorch_version='2.0',  
    py_version='py310',
    model_data="s3://sagemaker-us-east-1-811828458885/huggingface-qlora-2308180454/output/model.tar.gz", OR #huggingface_estimator.model_data
    role=role,
)

_response = huggingface_model.register(
    content_types=["application/json"],
    response_types=["application/json"],
    inference_instances=[
    "ml.p2.16xlarge",
    ...
    ],
    transform_instances=[
    "ml.p2.16xlarge",
    ...
    ],
    model_package_group_name=ft_model_pkg_group_name,
    approval_status="Approved"
)

>>>
Model collection creation status: {'added_groups': ['arn:aws:sagemaker:us-east-1:811828458885:model-package-group/mikael110--llama-2-7b-guanaco-fp16-finetuned-sql'], 'failure': []}

The following snippet shows a view from the Model Registry where the models are split into base and fine-tuned weights.

Managing models, datasets, and tasks for hyper-personalized LLMs can quickly become overwhelming. SageMaker Model Registry Collections can help you group related models together and organize them in a hierarchy to improve model discoverability. This makes it easier to track the relationships between base weights, adapter weights, and fine-tuning task datasets. You can also create complex relationships and linkages between models.

Create a new Collection and add your base model weights to this Collection

# create model collection
base_collection = model_collector.create(
    collection_name=model_group_for_base # ex: "Website_Customer_QnA_Bot_Model"
)

# Add the base weights at first level of model collections as all future models 
# are going to be tuned from the base weights
_response = model_collector.add_model_groups(
    collection_name=base_collection["Arn"],
    model_groups=[base_model_pkg_group_name]
)
print(f"Model collection creation status: {_response}")

>>>
Model collection creation status: {'added_groups': ['arn:aws:sagemaker:us-west-2:376678947624:model-package-group/Mikael110--llama-2-7b-guanaco-fp16'], 'failure': []}

Link all your Fine-Tuned LoRA Adapter Delta Weights to this collection by task and/or dataset

# create model collection for finetuned and link it back to the base
finetuned_collection = model_collector.create(
    collection_name=model_group_for_finetune,
    parent_collection_name=model_group_for_base
)

# add finetuned model package group to the new finetuned collection
_response = model_collector.add_model_groups(
    collection_name=model_group_for_finetune,
    model_groups=[ft_model_pkg_group_name]
)
print(f"Model collection creation status: {_response}")

>>>
Model collection creation status: {'added_groups': ['arn:aws:sagemaker:us-east-1:811828458885:model-package-group/mikael110--llama-2-7b-guanaco-fp16-finetuned-sql'], 'failure': []}

This will result in a collection hierarchy which are linked by model/task type and the dataset used to fine tune the base model.

This method of separating the base and adapter models has some drawbacks. One drawback is complexity in deploying the model. Because there are two separate model artifacts, you need additional steps to repackage the model instead of deploy directly from Model Registry. In the following code example, download and repack the latest version of the base model first.

!aws s3 cp {base_model_package.model_data} .

!tar -xvf {model_tar_filename} -C ./deepspeed/

!mv ./deepspeed/{model_id} ./deepspeed/base

!rm -rf ./deepspeed/{model_id}

Then download and repack the latest fine-tuned LoRA adapter weights.

!aws s3 cp {LoRA_package.model_data} .

!mkdir -p ./deepspeed/lora/

!tar -xzf model.tar.gz -C ./deepspeed/lora/

Since you will be using DJL serving with deepspeed to host the model, your inference directory should look like the following.

deepspeed
    |-serving.properties
    |-requirements.txt
    |-model.py
    |-base/
        |-...
    |-lora/
        |-...

Finally, package the custom inference code, base model, and LoRA adaptor in a single .tar.gz file for deployment.

!rm -f model.tar.gz
!tar czvf model.tar.gz -C deepspeed .
s3_code_artifact_deepspeed = sagemaker_session.upload_data("model.tar.gz", default_bucket, f"{s3_key_prefix}/inference")
print(f"S3 Code or Model tar for deepspeed uploaded to --- > {s3_code_artifact_deepspeed}")

Clean up

Clean up your resources by following the instructions in the cleanup section of the notebook. Refer to Amazon SageMaker Pricing for details on the cost of the inference instances.

Conclusion

This post walked you through best practices for managing LoRA fine-tuned models on Amazon SageMaker. We covered two main methods: combining the base and adapter weights into one self-contained model, and separating the base and adapter weights. Both approaches have tradeoffs, but separating weights helps optimize storage and enables advanced model management techniques like SageMaker Model Registry Collections. This allows you to build hierarchies and relationships between models to improve organization and discoverability. We encourage you to try the sample code on GitHub repository to experiment with these methods yourself. As generative AI progresses rapidly, following model management best practices will help you track experiments, find the right model for your task, and manage specialized LLMs efficiently at scale.

References

• Next generation Amazon SageMaker Experiments
• HuggingFace Callbacks

About the authors

James Wu is a Senior AI/ML Specialist Solution Architect at AWS. helping customers design and build AI/ML solutions. James’s work covers a wide range of ML use cases, with a primary interest in computer vision, deep learning, and scaling ML across the enterprise. Prior to joining AWS, James was an architect, developer, and technology leader for over 10 years, including 6 years in engineering and 4 years in marketing & advertising industries.

Pranav Murthy is an AI/ML Specialist Solutions Architect at AWS. He focuses on helping customers build, train, deploy and migrate machine learning (ML) workloads to SageMaker. He previously worked in the semiconductor industry developing large computer vision (CV) and natural language processing (NLP) models to improve semiconductor processes. In his free time, he enjoys playing chess and traveling.

Mecit Gungor is an AI/ML Specialist Solution Architect at AWS helping customers design and build AI/ML solutions at scale. He covers a wide range of AI/ML use cases for Telecommunication customers and currently focuses on Generative AI, LLMs, and training and inference optimization. He can often be found hiking in the wilderness or playing board games with his friends in his free time.

Shelbee Eigenbrode is a Principal AI and Machine Learning Specialist Solutions Architect at Amazon Web Services (AWS). She has been in technology for 24 years spanning multiple industries, technologies, and roles. She is currently focusing on combining her DevOps and ML background into the domain of MLOps to help customers deliver and manage ML workloads at scale. With over 35 patents granted across various technology domains, she has a passion for continuous innovation and using data to drive business outcomes. Shelbee is a co-creator and instructor of the Practical Data Science specialization on Coursera. She is also the Co-Director of Women In Big Data (WiBD), Denver chapter. In her spare time, she likes to spend time with her family, friends, and overactive dogs.

Read More

At a basic level, Machine Learning (ML) technology learns from data to make predictions. Businesses use their data with an ML-powered personalization service to elevate their customer experience. This approach allows businesses to use data to derive actionable insights and help grow their revenue and brand loyalty.

Amazon Personalize accelerates your digital transformation with ML, making it easier to integrate personalized recommendations into existing websites, applications, email marketing systems, and more. Amazon Personalize enables developers to quickly implement a customized personalization engine, without requiring ML expertise. Amazon Personalize provisions the necessary infrastructure and manages the entire machine learning (ML) pipeline, including processing the data, identifying features, using the most appropriate algorithms, and training, optimizing, and hosting the models. You receive results through an API and pay only for what you use, with no minimum fees or upfront commitments.

The post Architecting near real-time personalized recommendations with Amazon Personalize shows how to architect near real-time personalized recommendations using Amazon Personalize and AWS purpose-built data services. In this post, we walk you through a reference implementation of a real-time personalized recommendation system using Amazon Personalize.

Solution overview

The real-time personalized recommendations solution is implemented using Amazon Personalize, Amazon Simple Storage Service (Amazon S3), Amazon Kinesis Data Streams, AWS Lambda, and Amazon API Gateway.

The architecture is implemented as follows:

1. Data preparation – Start by creating a dataset group, schemas, and datasets representing your items, interactions, and user data.
2. Train the model – After importing your data, select the recipe matching your use case, and then create a solution to train a model by creating a solution version. When your solution version is ready, you can create a campaign for your solution version.
3. Get near real-time recommendations – When you have a campaign, you can integrate calls to the campaign in your application. This is where calls to the GetRecommendations or GetPersonalizedRanking APIs are made to request near real-time recommendations from Amazon Personalize.

For more information, refer to Architecting near real-time personalized recommendations with Amazon Personalize.

The following diagram illustrates the solution architecture.

Implementation

We demonstrate this implementation with a use case about making real-time movie recommendations to an end user based on their interactions with the movie database over time.

The solution is implemented using the following steps:

1. Prerequisite (Data preparation)
2. Setup your development environment
3. Deploy the solution
4. Create a solution version
5. Create a campaign
6. Create an event tracker
7. Get recommendations
8. Ingest real-time interactions
9. Validate real-time recommendations
10. Cleanup

Prerequisites

Before you get started, make sure you have the following prerequisites:

• Prepare your training data – Prepare and upload the data to an S3 bucket using the instructions. For this particular use case, you will be uploading interactions data and items data. An interaction is an event that you record and then import as training data. Amazon Personalize generates recommendations primarily based on the interactions data you import into an Interactions dataset. You can record multiple event types, such as click, watch, or like. Although the model created by Amazon Personalize can suggest based on a user’s past interactions, the quality of these suggestions can be enhanced when the model possesses data about the associations among users or items . If a user has engaged with movies categorized as Drama in the item dataset, Amazon Personalize will suggest movies (items) with the same genre.
• Setup your development environment – Install the AWS Command Line Interface (AWS CLI).
• Configure CLI with your Amazon account – Configure the AWS CLI with your AWS account information.
• Install and bootstrap AWS Cloud Development Kit (AWS CDK)

Deploy the solution

To deploy the solution, do the following:

• Clone the repository to a new folder on your desktop.
• Deploy the stack to your AWS environment.

Create a solution version

A solution refers to the combination of an Amazon Personalize recipe, customized parameters, and one or more solution versions (trained models). When you deploy the CDK project in the previous step, a solution with a User-Personalization recipe is created for you automatically. A solution version refers to a trained machine learning model. Create a solution version for the implementation.

Create a campaign

A campaign deploys a solution version (trained model) with a provisioned transaction capacity for generating real-time recommendations. Create a campaign for the implementation.

Create an event tracker

Amazon Personalize can make recommendations based on real-time event data only, historical event data only, or both. Record real-time events to build out your interactions data and allow Amazon Personalize to learn from your user’s most recent activity. This keeps your data fresh and improves the relevance of Amazon Personalize recommendations. Before you can record events, you must create an event tracker. An event tracker directs new event data to the Interactions dataset in your dataset group. Create and event tracker for the implementation.

Get recommendations

In this use case, the interaction dataset is composed of movie IDs. Consequently, the recommendations presented to the user will consist of movie IDs that align most closely with their personal preferences, determined from their historical interactions. You can use the getRecommendations API to retrieve personalized recommendations for a user by sending its associated userID, the number of results for recommendations that you need for the user as well as the campaign ARN. You can find the campaign ARN in the Amazon Personalize console menu.

For example, the following request will retrieve 5 recommendations for the user whose userId is 429:

curl --location 'https://{your-api-id}.execute-api.{your-region}.amazonaws.com/prod/getRecommendations?campaignArn={campaignArn}&userId=429&numResults=5'

The response from the request will be:

{
	"$metadata": {
		"httpStatusCode": 200,
		"requestId": "7159c128-4e16-45a4-9d7e-cf19aa2256e8",
		"attempts": 1,
		"totalRetryDelay": 0
	},
	"itemList": [
	{
		"itemId": "596",
		"score": 0.0243044
	},
	{
		"itemId": "153",
		"score": 0.0151695
	},
	{
		"itemId": "16",
		"score": 0.013694
	},
	{
		"itemId": "261",
		"score": 0.013524
	},
	{
		"itemId": "34",
		"score": 0.0122294
	}
	],
	"recommendationId": "RID-1d-40c1-8d20-dfffbd7b0ac7-CID-06b10f"
}

The items returned by the API call are the movies that Amazon Personalize recommends to the user based on their historical interactions.

The score values provided in this context represent floating-point numbers that range between zero and 1.0. These values correspond to the current campaign and the associated recipes for this use case. They are determined based on the collective scores assigned to all items present in your comprehensive dataset.

Ingest real-time interactions

In the previous example, recommendations were obtained for the user with an ID of 429 based on their historical interactions with the movie database. For real-time recommendations, the user interactions with the items must be ingested into Amazon Personalize in real-time. These interactions are ingested into the recommendation system through the Amazon Personalize Event Tracker. The type of interaction, also called EventType, is given by the column of the same name in the interaction data dataset (EVENT_TYPE). In this example, the events can be of type “watch” or “click”, but you can have your own types of events according to the needs of your application.

In this example, the exposed API that generates the events of the users with the items receives the “interactions” parameter that corresponds to the number of events (interactions) of a user (UserId) with a single element (itemId) right now. The trackingId parameter can be found in the Amazon Personalize console and in the response of the creation of Event Tracker request.

This example shows a putEvent request: Generate 1 interactions of click type, with an item id of ‘185’ for the user id ‘429’, using the current timestamp. Note that in production, the ‘sentAt’ should be set to the time of the user’s interaction. In the following example, we set this to the point in time in epoch time format when we wrote the API request for this post. The events are sent to Amazon Kinesis Data Streams through an API Gateway which is why you need to send the stream-name and PartitionKey parameters.

curl --location 'https://iyxhva3ll6.execute-api.us-west-2.amazonaws.com/prod/data' --header 'Content-Type: application/json' --data '{ "stream-name": "my-stream","Data": {"userId" : "429", "interactions": 1, "itemId": "185", "trackingId" : "c90ac6d7-3d89-4abc-8a70-9b09c295cbcd", "eventType": "click", "sentAt":"1698711110"},"PartitionKey":"userId"}'

You will receive a confirmation response similar to the following:

{
	"Message": "Event sent successfully",
	"data": {
		"EncryptionType": "KMS",
		"SequenceNumber": "49..........1901314",
		"ShardId": "shardId-xxxxxxx"
	}
}

Validate real-time recommendations

Because the interaction dataset has been updated, the recommendations will be automatically updated to consider the new interactions. To validate the recommendations updated in real-time, you can call the getRecommendations API again for the same user id 429, and the result should be different from the previous one. The following results show a new recommendation with an id of 594 and the recommendations with the id’s of 16, 596, 153and 261 changed their scores. These items brought in new movie genre (‘Animation|Children|Drama|Fantasy|Musical’) the top 5 recommendations.

Request:

curl --location 'https://{your-api-id}.execute-api.{your-region}.amazonaws.com/prod/getRecommendations?campaignArn={campaignArn} &userId=429&numResults=5'

Response:

{
	"$metadata": {
		"httpStatusCode": 200,
		"requestId": "680f2be8-2e64-47d7-96f7-1c4aa9b9ac9d",
		"attempts": 1,
		"totalRetryDelay": 0
	},
	"itemList": [
	{
		"itemId": "596",
		"score": 0.0288085
	},
	{
		"itemId": "16",
		"score": 0.0134173
	},
	{
		"itemId": "594",
		"score": 0.0129357
	},
	{
		"itemId": "153",
		"score": 0.0129337
	},
	{
		"itemId": "261",
		"score": 0.0123728
	}
	],
	"recommendationId": "RID-dc-44f8-a327-482fb9e54921-CID-06b10f"
}

The response shows that the recommendation provided by Amazon Personalize was updated in real-time.

Clean up

To avoid unnecessary charges, clean up the solution implementation by using Cleaning up resources.

Conclusion

In this post, we showed you how to implement a real-time personalized recommendations system using Amazon Personalize. The interactions with Amazon Personalize to ingest real-time interactions and get recommendations were executed through a command line tool called curl but these API calls can be integrated into a business application and derive the same outcome.

To choose a new recipe for your use case, refer to Real-time personalization. To measure the impact of the recommendations made by Amazon Personalize, refer to Measuring impact of recommendations.

About the Authors

Cristian Marquez is a Senior Cloud Application Architect. He has vast experience designing, building, and delivering enterprise-level software, high load and distributed systems and cloud native applications. He has experience in backend and frontend programming languages, as well as system design and implementation of DevOps practices. He actively assists customers build and secure innovative cloud solutions, solving their business problems and achieving their business goals.

Anand Komandooru is a Senior Cloud Architect at AWS. He joined AWS Professional Services organization in 2021 and helps customers build cloud-native applications on AWS cloud. He has over 20 years of experience building software and his favorite Amazon leadership principle is “Leaders are right a lot.“

Read More

One of the most common applications of generative AI and large language models (LLMs) is answering questions based on a specific external knowledge corpus. Retrieval-Augmented Generation (RAG) is a popular technique for building question answering systems that use an external knowledge base. To learn more, refer to Build a powerful question answering bot with Amazon SageMaker, Amazon OpenSearch Service, Streamlit, and LangChain.

Traditional RAG systems often struggle to provide satisfactory answers when users ask vague or ambiguous questions without providing sufficient context. This leads to unhelpful responses like “I don’t know” or incorrect, made-up answers provided by an LLM. In this post, we demonstrate a solution to improve the quality of answers in such use cases over traditional RAG systems by introducing an interactive clarification component using LangChain.

The key idea is to enable the RAG system to engage in a conversational dialogue with the user when the initial question is unclear. By asking clarifying questions, prompting the user for more details, and incorporating the new contextual information, the RAG system can gather the necessary context to provide an accurate, helpful answer—even from an ambiguous initial user query.

Solution overview

To demonstrate our solution, we have set up an Amazon Kendra index (composed of the AWS online documentation for Amazon Kendra, Amazon Lex, and Amazon SageMaker), a LangChain agent with an Amazon Bedrock LLM, and a straightforward Streamlit user interface.

Prerequisites

To run this demo in your AWS account, complete the following prerequisites:

1. Clone the GitHub repository and follow the steps explained in the README.
2. Deploy an Amazon Kendra index in your AWS account. You can use the following AWS CloudFormation template to create a new index or use an already running index. Deploying a new index might add additional charges to your bill, therefore we recommend deleting it if you don’t longer need it. Note that the data within the index will be sent to the selected Amazon Bedrock foundation model (FM).
3. The LangChain agent relies on FMs available in Amazon Bedrock, but this can be adapted to any other LLM that LangChain supports.
4. To experiment with the sample front end shared with the code, you can use Amazon SageMaker Studio to run a local deployment of the Streamlit app. Note that running this demo will incur some additional costs.

Implement the solution

Traditional RAG agents are often designed as follows. The agent has access to a tool that is used to retrieve documents relevant to a user query. The retrieved documents are then inserted into the LLM prompt, so that the agent can provide an answer based on the retrieved document snippets.

In this post, we implement an agent that has access to KendraRetrievalTool and derives relevant documents from the Amazon Kendra index and provides the answer given the retrieved context:

# tool for Kendra retrieval

kendra_tool = Tool(
    name="KendraRetrievalTool",
    func=retrieval_qa_chain,
    description="Use this tool first to answer human questions. The input to this tool should be the question.",
)
# traditional RAG agent

traditional_agent = initialize_agent(
    agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
    tools=[kendra_tool]
    llm=llm,
    early_stopping_method="generate",
    memory=conversational_memory,
)

# user question
answer = traditional_agent.run("How many GPUs does my EC2 instance have?")

Refer to the GitHub repo for the full implementation code. To learn more about traditional RAG use cases, refer to Question answering using Retrieval Augmented Generation with foundation models in Amazon SageMaker JumpStart.

Consider the following example. A user asks “How many GPUs does my EC2 instance have?” As shown in the following screenshot, the agent is looking for the answer using KendraRetrievalTool. However, the agent realizes it doesn’t know which Amazon Elastic Compute Cloud (Amazon EC2) instance type the user is referencing and therefore provides no helpful answer to the user, leading to a poor customer experience.

To address this problem, we define an additional custom tool called AskHumanTool and provide it to the agent. The tool instructs an LLM to read the user question and ask a follow-up question to the user if KendraRetrievalTool is not able to return a good answer. This implies that the agent will now have two tools at its disposal:

# tool for asking human
human_ask_tool = CustomAskHumanTool()

# RAG agent with two tools
improved_agent = initialize_agent(
    agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
    tools=[kendra_tool, human_ask_tool],
    llm=llm,
    early_stopping_method="generate",
    memory=conversational_memory,
)

# user question
answer = improved_agent.run("How many GPUs does my EC2 instance have?")

This allows the agent to either refine the question or provide additional context that is needed to respond to the prompt. To guide the agent to use AskHumanTool for this purpose, we provide the following tool description to the LLM:

Use this tool if you don’t find an answer using the KendraRetrievalTool. 
Ask the human to clarify the question or provide the missing information. 
The input should be a question for the human.

As illustrated in the following screenshot, by using AskHumanTool, the agent is now identifying vague user questions and returning a follow-up question to the user asking to specify what EC2 instance type is being used.

After the user has specified the instance type, the agent is incorporating the additional answer into the context for the original question, before deriving the correct answer.

Note that the agent can now decide whether to use KendraRetrievalTool to retrieve the relevant documents or ask a clarifying question using AskHumanTool. The agent’s decision is based on whether it finds the document snippets inserted into the prompt sufficient to provide the final answer. This flexibility allows the RAG system to support different queries a user may submit, including both well-formulated and vague questions.

In our example, the full agent workflow is as follows:

1. The user makes a request to the RAG app, asking “How many GPUs does my EC2 instance have?”
2. The agent uses the LLM to decide what action to take: Find relevant information to answer the user’s request by calling the KendraRetrievalTool.
3. The agent retrieves information from the Amazon Kendra index using the tool. The snippets from the retrieved documents are inserted into the agent prompt.
4. The LLM (of the agent) derives that the retrieved documents from Amazon Kendra aren’t helpful or don’t contain enough context to provide an answer to the user’s request.
5. The agent uses AskHumanTool to formulate a follow-up question: “What is the specific EC2 instance type you are using? Knowing the instance type would help answer how many GPUs it has.” The user provides the answer “ml.g5.12xlarge,” and the agent calls KendraRetrievalTool again, but this time adding the EC2 instance type into the search query.
6. After running through Steps 2–4 again, the agent derives a useful answer and sends it back to the user.

The following diagram illustrates this workflow.

The example described in this post illustrates how the addition of the custom AskHumanTool allows the agent to request clarifying details when needed. This can improve the reliability and accuracy of the responses, leading to a better customer experience in a growing number of RAG applications across different domains.

Clean up

To avoid incurring unnecessary costs, delete the Amazon Kendra index if you’re not using it anymore and shut down the SageMaker Studio instance if you used it to run the demo.

Conclusion

In this post, we showed how to enable a better customer experience for the users of a RAG system by adding a custom tool that enables the system to ask a user for a missing piece of information. This interactive conversational approach represents a promising direction for improving traditional RAG architectures. The ability to resolve vagueness through a dialogue can lead to delivering more satisfactory answers from a knowledge base.

Note that this approach is not limited to RAG use cases; you can use it in other generative AI use cases that depend on an agent at its core, where a custom AskHumanTool can be added.

For more information about using Amazon Kendra with generative AI, refer to Quickly build high-accuracy Generative AI applications on enterprise data using Amazon Kendra, LangChain, and large language models.

About the authors

Antonia Wiebeler is a Data Scientist at the AWS Generative AI Innovation Center, where she enjoys building proofs of concept for customers. Her passion is exploring how generative AI can solve real-world problems and create value for customers. While she is not coding, she enjoys running and competing in triathlons.

Nikita Kozodoi is an Applied Scientist at the AWS Generative AI Innovation Center, where he develops ML solutions to solve customer problems across industries. In his role, he focuses on advancing generative AI to tackle real-world challenges. In his spare time, he loves playing beach volleyball and reading.

Read More