MT-Bench (GPT-4-Turbo judge, corrected references)
Chat & Writing · 2024-06-12
NVIDIA's corrected variant of LMSYS MT-Bench: the 80 two-turn questions (160 turns) across writing, roleplay, extraction, reasoning, math, coding, STEM and humanities are answered by the model and graded 1 to 10 by GPT-4-0125-Preview instead of GPT-4, with the reasoning, math and coding reference answers corrected (FastChat pull request 3158). The score is the mean over all turns; NVIDIA reports it runs about 0.8 point below original GPT-4-judged MT-Bench, so the two are kept separate. Higher is better.