Atlas

Benchmarks

← All benchmarks

MedQA 4-options (Open Medical-LLM)

Science · 2024-04-19

MedQA draws US Medical Licensing Examination style questions. This is the four-option variant's 1,273-question test split, scored by the Open Medical-LLM Leaderboard under lm-evaluation-harness with a fixed prompt. That protocol differs from lab-reported MedQA runs and from the Vals demographic-bias variants Atlas already carries, so it forms its own factor group.

Top models (higher is better)

ModelScore
GPT-478.8
Llama 3 8B Instruct60.7
Llama 3 8B59.7
GPT-3.5 Turbo 110657.7
Solar 10.7B Instruct V1.054.7
Zephyr 7B Beta51.5
Mistral 7B v0.1 Base50.8
Qwen1.5 7B48.9
Gemma 7B48.7
Gemma 1.1 7B IT45.7
Qwen1.5 7B Chat45.4
Mistral 7B Instruct v0.143.5
Gemma 7B IT40.4
Phi-1.531.1
Gemma 2B29.2
Falcon 7B27.9
Loading Atlas data…