Skip to content

Benchmark

WildBench: haiku_reward.task_macro.K=-1

People compare model answers and choose which they prefer. Higher scores are better.

Models measured
32
Scores recorded
32
Sources
1
IQ weight
1.26

Score distribution

2.3
Mean
7.6
Median
33.8
Std dev
-32.5
10th percentile
46.2
90th percentile
-76.1
Min
-76.152.6
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.26 for chat models, measured over 32 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1GPT 4oallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data52.6
2Claude 3.5 Sonnetallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data48.6
3Gemini 1.5 Proallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data48.3
4GPT-4 Turboallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data46.2
5GPT-4-0125-previewallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data38.5
6Claude 3 Opusallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data36.4
7Gemini 1.5 Flashallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data34.6
8Llama 3 70B Instructallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data32.2
9Nemotron-4 340B Instructallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data31.2
10yi-largeallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data28.2
11Reka Coreallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data27.1
12Yi 1.5 34B Chatallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data23.5
13Claude 3 Sonnetallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data21.6
14DeepSeek Coder V2 Instructallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data19.7
15Qwen2 72B Instructallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data12.5
16GLM-4-9b-chatallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data7.6
17Command R Plusallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data1.5
18Mistral Largeallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307, 2402· WildBench direct data-8.5
19Starling LM 7B Betaallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-10.6
20Reka Flashallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-10.8
21Llama 3 8B Instructallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-14.6
22Mixtral 8x7B Instruct v0.1allenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-18.1
23Nous Hermes 2 Mixtral 8x7B DPOallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-19.7
24Command R+allenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-19.8
25Tulu 2 DPO 70Ballenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-25.9
26Reka Edgeallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-26.5
27Mistral 7B Instruct v0.2allenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-29.2
28GPT-3.5 Turboallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-32.1
29Llama-2-70b-chat-hfallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-32.5
30Llama-2-7b-chat-hfallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-52.4
31Gemma 7B ITallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-59
32Gemma 2B ITallenai/WildBench, haiku_reward.task_macro.K: -1, claude-3-haiku-20240307· WildBench direct data-76.1

Data credits: WildBench, Allen Institute for AI