Preference
MT-Bench
MT-Bench rates how well chat models handle open ended questions asked over two conversational turns, with a strong model acting as the judge and scoring each answer on its own.
160items
34subjects
CC-BY-4.0license
preferencedomain
textmodality
item-level responses released
Saturation status: No
Response matrix
Loading session strips…
lowhighUnobserved
Scale: {1, 2, ..., 10}