LLM-as-judge
Using a language model to grade another model's output against a rubric, for qualities that code cannot check.
In more detail
Zheng and colleagues showed in 2023 that strong judges agree with people about as often as people agree with each other, but have biases. Give the judge a specific rubric, ask it to reason before ruling and allow "unknown", and measure it against human labels with kappa before trusting it.
Related terms
- rubric — The written criteria a grader applies to decide whether an output passes, ideally specific, observable and with a way out for "unknown".
- Cohen's kappa — A measure of agreement between two raters that corrects for the agreement expected by chance, from 1 for perfect to 0 for chance level.
- judge bias — Systematic errors of model graders, such as favouring the first answer shown, longer answers or their own model's style.
- eval harness — The code that runs every eval case through the system, applies the checks and reports scores per case and overall.
Taught in
- What An AI Engineer Does — AI Engineering: LLM Apps, RAG & Agents
- What An Embedding Is — AI Engineering: LLM Apps, RAG & Agents
- Vector Indexes And Databases — AI Engineering: LLM Apps, RAG & Agents
- Keyword And Hybrid Search — AI Engineering: LLM Apps, RAG & Agents
- Retrieval That Finds The Right Thing — AI Engineering: LLM Apps, RAG & Agents
- Eval-Driven Development — AI Engineering: LLM Apps, RAG & Agents
- Model-Graded Evals — AI Engineering: LLM Apps, RAG & Agents
- Evaluating RAG And Agents — AI Engineering: LLM Apps, RAG & Agents
- Regression Tests In CI — AI Engineering: LLM Apps, RAG & Agents
- Images, Audio And Speech — AI Engineering: LLM Apps, RAG & Agents
- Capstone: The Tasks Assistant — AI Engineering: LLM Apps, RAG & Agents
References
- Zheng et al. (2023): Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
- OpenAI: Graders (score model graders)