Atlas

Benchmarks

← All benchmarks

RACE-High

General QA · 2017-04-15

RACE-High is the high-school portion of RACE, a four-option multiple-choice reading-comprehension dataset built from English exams for Chinese students. Its passages and questions are harder than RACE-Middle; scores are accuracy on its 3,498-question test split.

Top models (higher is better)

ModelScore
Opus 392.9
Claude 1.388.8
Claude 3 Sonnet88.8
Claude 288.3
Claude 3 Haiku87.0
Claude Instant 1.185.5
Llama 3 70B57.9
Llama 3.1 405B56.8
Mixtral 8x22B56.7
DeepSeek LLM 67B Chat56.0
LLaMA 65B51.6
DeepSeek-V3-Base51.3
DeepSeek LLM 7B Chat50.8
DeepSeek LLM 67B Base50.7
DeepSeekMoE 16B Chat50.6
Qwen2.5 72B50.3
PaLM 540B49.1
LLaMA 33B48.3
PaLM 62B47.5
LLaMA-13B47.2
Qwen1.5 72B47.0
LLaMA 7B46.9
DeepSeek LLM 7B Base46.5
DeepSeekMoE 16B Base46.4
Loading Atlas data…