Skip to content

Benchmark

WildBench: haiku_reward.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
1.36

Score distribution

4.7
Mean
9.5
Median
28.6
Std dev
-27.4
10th percentile
39.1
90th percentile
-69.6
Min
-69.643.4
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.36 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-4 Turboallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data43.4
2GPT 4oallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data43.3
3Claude 3.5 Sonnetallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data40.9
4Gemini 1.5 Proallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data39.1
5GPT-4-0125-previewallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data37.8
6Claude 3 Opusallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data32
7Llama 3 70B Instructallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data30.1
8Nemotron-4 340B Instructallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data28.9
9Gemini 1.5 Flashallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data27.1
10yi-largeallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data26.6
11Reka Coreallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data26.3
12Yi 1.5 34B Chatallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data24.8
13DeepSeek Coder V2 Instructallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data20.2
14Claude 3 Sonnetallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data18.4
15Qwen2 72B Instructallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data13
16GLM-4-9b-chatallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data9.5
17Command R Plusallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data7.4
18Mistral Largeallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307, 2402· WildBench direct data-3
19Starling LM 7B Betaallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-3.2
20Reka Flashallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-5.2
21Llama 3 8B Instructallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-7.3
22Command R+allenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-10.5
23Mixtral 8x7B Instruct v0.1allenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-11.6
24Nous Hermes 2 Mixtral 8x7B DPOallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-16.3
25Reka Edgeallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-17
26Tulu 2 DPO 70Ballenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-17
27Mistral 7B Instruct v0.2allenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-19.3
28Llama-2-70b-chat-hfallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-20
29GPT-3.5 Turboallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-27.4
30Llama-2-7b-chat-hfallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-39.1
31Gemma 7B ITallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-52.4
32Gemma 2B ITallenai/WildBench, haiku_reward.K: -1, claude-3-haiku-20240307· WildBench direct data-69.6

Data credits: WildBench, Allen Institute for AI