Get started with model evaluations

A large language model (LLM) is a machine learning model that can analyze and generate natural language text. If you want to evaluate an LLM, SageMaker provides the following three options that you can choose:

You can either use an algorithm to automatically evaluate your foundation model or ask a human work team to evaluate the models' responses.

Human work teams can evaluate and compare up to two models concurrently using metrics that indicate preference for one response over another. The work flow, metrics, and instructions for a human evaluation can be tailored to fit a particular use case. Humans can also provide a more refined evaluation than an algorithmic evaluation.

You can also use an algorithm to evaluate your LLM using benchmarks to rapidly score your model responses in Studio. Studio provides a guided work flow to evaluate responses from a JumpStart model using pre-defined metrics. These metrics are specific to generative AI tasks. This guided flow uses built-in or custom datasets to evaluate your LLM.

Alternatively, you can use the fmeval library to create a more customized work flow using automatic evaluations than what is available in Studio. Using Python code and the fmeval library, you can evaluate any text-based LLM, including models that were created outside of JumpStart.

The following topics provide an overview of foundation model evaluations, a summary of the automatic and human Foundation Model Evaluation (FMEval) work flows, how to run them, and how to view an analysis report of your results. The automatic evaluation topic shows how to configure and run both a starting and customized evaluation.

Topics