Evaluation is the process of continuously improving your LLM application. This requires a way to judge your application’s outputs, which are often natural language. Using an LLM to grade natural language outputs (e.g., for correctness relative to a reference answer, tone, or conciseness) is a popular approach, but requires prompt engineering and careful auditing of the LLM judge!
Our new release of LangSmith presents a solution to this rising problem, allowing a user to (1) correct LLM-as-a-Judge outputs and then (2) pass those corrections back to the judge as few-shot example for future iterations. This creates LLM-as-a-Judge evaluators grounded in human feedback that better encode your preferences without the need for challenging prompt engineering.
Here we show how apply Corrections + Few Shot to online evaluators that are pinned to a dataset.
4 окт 2024