Preference
MT-Bench
MT-Bench GPT-4 single-answer judgment scores (34 models, 80 questions, 2 turns).
160items
34subjects
CC-BY-4.0license
preferencedomain
textmodality
item-level responses released
Saturation status: Unknown
Response matrix
Fit to width. Hover for subject & item; click a cell for details.
lowhighUnobserved
Scale: {1, 2, ..., 10}
Subjects
- 1gpt-48.9906
- 2gpt-3.5-turbo7.9437
- 3claude-instant-v17.9057
- 4claude-v17.9
- 5vicuna-33b-v1.37.1219
- 6wizardlm-30b7.0094
- 7Llama-2-70b-chat6.8563
- 8Llama-2-13b-chat6.65
- 9guanaco-33b6.5281
- 10tulu-30b6.4344
- 11guanaco-65b6.4094
- 12oasst-sft-7-llama-30b6.4094
- 13palm-2-chat-bison-0016.4
- 14mpt-30b-chat6.3937
- 15vicuna-13b-v1.36.3875
- 16wizardlm-13b6.3531
- 17Llama-2-7b-chat6.2687
- 18vicuna-7b-v1.35.9969
- 19baize-v2-13b5.75
- 20nous-hermes-13b5.5535
- 21mpt-7b-chat5.4591
- 22gpt4all-13b-snoozy5.4528
- 23koala-13b5.35
- 24mpt-30b-instruct5.2188
- 25falcon-40b-instruct5.1688
- 26h2ogpt-oasst-open-llama-13b4.625
- 27alpaca-13b4.5312
- 28chatglm-6b4.5
- 29oasst-sft-4-pythia-12b4.3187
- 30rwkv-4-raven-14b3.9844
- 31dolly-v2-12b3.275
- 32fastchat-t5-3b3.0406
- 33stablelm-tuned-alpha-7b2.7531
- 34llama-13b2.6063