Atlas

Benchmarks

← All benchmarks

AGIEval

General QA · 2023-04-13

AGIEval in its full form pools the English and Chinese tasks of AGIEval, a benchmark built from standardized admission and qualification exams such as the Gaokao, SAT, LSAT, law and civil-service exams. Scores are accuracy averaged over the tasks; the English-only and Chinese-only portions are separate rows. Higher is better.

Top models (higher is better)

ModelScore
DeepSeek-V3-Base79.6
Qwen2.5 72B75.8
Qwen1.5 72B64.4
Qwen1.5 72B Chat62.8
DeepSeek-V2-Chat61.4
Llama 3.1 405B60.6
DeepSeek-Coder-V2-Instruct60.0
Llama 3 70B Instruct56.6
Llama 3 70B49.8
DeepSeek LLM 67B Chat46.4
Mixtral 8x22B43.4
DeepSeek-V2-Lite-Chat42.8
Mixtral 8x22B Instruct41.4
DeepSeek LLM 67B Base41.3
DeepSeek-V2-Lite33.2
DeepSeek-Coder-V2-Lite-Instruct28.7
DeepSeek-VL 7B Chat27.8
DeepSeek LLM 7B Base26.4
DeepSeekMoE 16B Chat19.7
DeepSeekMoE 16B Base17.4
DeepSeek-VL 1.3B Chat14.0
Loading Atlas data…