An automatic evaluator for instruction-following language models. Human-validated, high-quality, cheap, and fast.
Terminal score: 0–1000 raw, weighted across 4 dimensions. Public score: 0–10 normalized (shown in the 30-day stars chart above).
