Get started quickly with curated datasets, such as CrowS-Pairs, TriviaQA, and WikiText, and curated algorithms, such as Bert-Score, Rouge, and F1. You can customize your own prompt datasets and scoring algorithms specific for your generative AI application. The automatic evaluation is also available as an open-source library in GitHub to allow you to run it anywhere. Sample notebooks show you how to programmatically run evaluation for any FMs, including models that are not hosted on AWS, and how to integrate FM evaluations with SageMaker MLOps and governance tools, such as SageMaker Pipelines, SageMaker Model Registry, and SageMaker Model Cards.

Some evaluation criteria are nuanced or subjective and require human judgement to assess. In addition to automated, metrics- based evaluations, you can ask humans (either your own employees or an AWS-managed evaluation team), to evaluate model outputs on dimensions like helpfulness, tone, and adherence to brand voice. Human evaluators can also check for consistency with company-specific guidelines, nomenclature, and brand voice. Setup custom instructions to instruct your evaluation team on how to evaluate prompts, for example by ranking or indicating thumbs up/down.

Evaluate your FM to determine if it provides high-quality responses for your specific generative AI task using automatic and/or human-based evaluations. Evaluate model accuracy with specific evaluation algorithms, such as Bert Score, Rouge and F1, tailored for specific generative AI tasks, such as summarization, question answering (Q&A), and classification. Check the semantic robustness of your FM output when prompted with semantic-preserving perturbations to the inputs, such as ButterFingers, random upper case, and whitespace add remove.

Evaluate the risk that your FM encoded stereotypes along the categories of race/color, gender/gender identity, sexual orientation, religion, age, nationality, disability, physical appearance, and socioeconomic status using automatic and/or human-based evaluations. You can also evaluate the risk of toxic content. These evaluations can be applied to any task that involves generation of content, including open-ended generation, summarization, and question answering.

Changes in live data can expose a new behavior of your model. For example, a credit risk prediction model trained on the data from one geographical region could change the importance it assigns to various features when applied to the data from another region. SageMaker Clarify is integrated with SageMaker Model Monitor to notify you using alerting systems such as CloudWatch if the importance of input features shift, causing model behavior to change.

After you’ve trained your model, you can run a SageMaker Clarify bias analysis through Amazon SageMaker Experiments to check your model for potential bias such as predictions that produce a negative result more frequently for one group than they do for another. You specify input features with respect to which you would like to measure bias in the model outcomes, and SageMaker runs an analysis and provides you with a visual report that identifies the different types of bias for each feature. AWS open-source method Fair Bayesian Optimization can help mitigate bias by tuning a model’s hyperparameters.

Bias can be introduced or exacerbated in deployed ML models when the training data differs from the live data that the model sees during deployment. For example, the outputs of a model for predicting home prices can become biased if the mortgage rates used to train the model differ from current mortgage rates. SageMaker Clarify bias detection capabilities are integrated into Amazon SageMaker Model Monitor so that when SageMaker detects bias beyond a certain threshold, it automatically generates metrics that you can view in Amazon SageMaker Studio and through Amazon CloudWatch metrics and alarms.