Atlas

Benchmarks

← All benchmarks

AlignBench (v1.0 or version unstated)

Chat & Writing · 2023-11-30

AlignBench is a Chinese alignment benchmark of user queries across eight categories, including reasoning, writing, role play and professional knowledge; a GPT-4 judge grades each answer on a 1-10 scale with reference answers and multi-dimensional rules, and the score is the average grade. This row holds results that do not name v1.1: the DeepSeek LLM, V2 and Coder-V2 and the Qwen2 reports predate the v1.1 release of June 15, 2024, so they are on v1.0, while later DeepSeek cards do not state the version. AlignBench v1.1 is the separate alignbench-v1-1. Higher is better.

Top models (higher is better)

ModelScore
Qwen2 72B Instruct8.3
DeepSeek-V2.58.0
DeepSeek-V2-Chat7.9
DeepSeek-Coder-V2-Instruct-07247.9
DeepSeek-V2-Chat-06287.9
Qwen1.5 110B Chat7.9
DeepSeek-Coder-V2-Instruct7.8
Llama 3 70B Instruct7.4
Qwen2-57B-A14B Instruct7.4
Qwen2 7B Instruct7.2
Yi 1.5 34B Chat7.2
Qwen1.5 32B Chat7.2
Yi 1.5 9B Chat6.9
DeepSeek-Coder-V2-Lite-Instruct6.8
DeepSeek LLM 67B Chat6.4
Llama 3 8B Instruct6.2
Qwen1.5 7B Chat6.2
Yi 34B Chat6.1
DeepSeek-V2-Lite-Chat6.0
Mixtral 8x7B Instruct5.7
Loading Atlas data…