Exam-style
Before a generative feature launches, the legal team wants a measured report on toxicity and bias across a 2,000-prompt test set, using a foundation model the company did not train. What produces that report?
Reveal the answer
A. The open-source fmeval library, scoring toxicity, prompt stereotyping, accuracy, and semantic robustness over the prompt set
fmeval is the engine behind SageMaker Clarify’s foundation model evaluation, and it installs on its own with pip. It scores generated text on toxicity, prompt stereotyping, accuracy, factual knowledge, and semantic robustness, which is the shape bias takes for a generative model. An Amazon Bedrock evaluation job is the managed alternative: its automatic metrics stop at accuracy, robustness, and toxicity, while its judge-based jobs add built-in Stereotyping and Harmfulness metrics. Guardrails apply content filters at request time, returning a per-category confidence band and a block decision for each call rather than a scored report over a prompt set. Model Cards record a judgement rather than measure one. The classical pre-training bias metrics need labelled training data the team does not have for someone else’s foundation model, and Comprehend reports sentiment per document or entity rather than stereotyping or quality disparity.
Q. You must measure a GenAI feature for bias and toxicity before launch. Which tool?
A. The open-source fmeval library, the engine behind SageMaker Clarify’s foundation model evaluation, scores accuracy, factual knowledge, toxicity, semantic robustness, and prompt stereotyping (bias) from your own notebook or pipeline. An Amazon Bedrock evaluation job is the managed path: its automatic metrics are accuracy, robustness, and toxicity, and its judge-based jobs add built-in Stereotyping and Harmfulness metrics. Clarify is closed to new customers and AWS plans no new features for it, though AWS puts no date on either; existing customers carry on, and pip install fmeval works either way.
Why? Bias for a generative model shows up as stereotyping and quality disparity, measured offline, not as label parity.