LLM-as-a-Judge

Under Development

This recipe is under active development. The page below sketches the planned scope. The launch script, judge prompts, and reference configs are not yet published.

---

What it is

LLM-as-a-Judge turns a separate language model into the reward function. For every rollout produced by the policy, a judge LLM reads the prompt together with the response (and optionally a rubric, reference answer, or multiple candidates) and emits a score, typically a scalar in [0, 1], a discrete rating, or a pairwise preference. That score is then used as the RL reward signal, so any algorithm in the RLVR family (PPO / GRPO / DAPO) can train against it.

This unlocks RL on tasks where no programmatic checker exists: instruction following, creative writing, dialogue, summarization, safety alignment, and so on.