Evaluating Generative AI Systems
Explore methods to evaluate generative AI systems, including automated metrics, human judgment, bias detection, and retrieval-augmented generation assessments. Understand how to apply AWS Bedrock and SageMaker AI tools for consistent, reliable, and responsible evaluation before deploying AI models.
A classifier has one right label, so accuracy settles the question. A generative model can answer the same question in a hundred acceptable ways, and it can also be fluent, confident, and wrong. “It looked good in the playground” is not evidence. Before a model, a prompt, or a knowledge base change goes to production, the team needs scores it can compare, human judgment where scores can’t capture quality, and checks for bias and unsupported claims.
This lesson covers the evaluation toolkit and how to choose the right method. Evaluation is done when every candidate (model, prompt, or retrieval setting) runs on the same representative dataset, the scores and the reviewer notes are stored, and the metric you chose can actually detect the failure you’re worried about.
The diagram below shows how the evaluation methods in this lesson fit together.
Reference-based NLP metrics
Reference-based metrics compare a model’s output with a human-written reference. They are cheap, repeatable, and need no judge, which makes them a good regression test. They differ in what “similar” means, as the following table shows.
Metric | What It Compares | Best Fit | Main Weakness |
BLEU | N-gram precision of the output against the reference, with a penalty for outputs that are too short | Machine translation | Rewards exact wording and ignores meaning |
ROUGE (ROUGE-N, ROUGE-L) | N-gram recall (ROUGE-N) or longest common subsequence (ROUGE-L) against the reference | Summarization | A correct paraphrase scores low |
BERTScore | Similarity of contextual token embeddings between output and reference | Summarization, question answering, any task where wording varies | Needs an embedding model, so it’s slower and less transparent |
Semantic similarity | Cosine similarity between the embedding of the whole output and the whole reference | Paraphrase detection, comparing answers to a reference answer | One number for the whole text, so a single wrong fact barely moves it |
Two memory aids help on the exam. BLEU is precision-oriented and associated with translation. ROUGE is recall-oriented and associated with summarization. When a scenario says the model paraphrases correctly but scores poorly, the metric is counting word overlap, so the better choice is an embedding-based metric such as BERTScore or semantic similarity.
Embedding-based scores still miss factual errors. A summary that reverses “approved” to “denied” is close to the reference in embedding space. Pair these metrics with a faithfulness or correctness check from a judge or a human.
Amazon Bedrock and SageMaker AI expose different subsets of these metrics. Bedrock automatic evaluation reports BERTScore for summarization and F1 for question answering. SageMaker AI Clarify foundation model evaluation, including the fmeval library, reports ROUGE-N, METEOR, and BERTScore for summarization. A metric that a managed job doesn’t offer, such as BLEU, is computed in your own code over the model outputs.
Amazon Bedrock evaluations
Amazon Bedrock evaluations run the dataset through the model, score the responses, and write a report to Amazon S3, so no separate evaluation pipeline is needed. ...