Atlas

Benchmarks

← All benchmarks

SciCode

Science · 2024-07-12

SciCode is a scientist-curated benchmark with 80 research coding problems decomposed into 338 subproblems; the standard test split has 288 subproblems from 65 main problems. This definition represents the official standard setup without scientist-provided background knowledge or reasoning steps. It measures whether models can generate executable scientific code from research-style specifications; background-assisted and modified-library evaluations require separate protocol identities.

Top models (higher is better)

ModelScore
Beam 1.0 501B A23B49.2
o3-mini34.4
DeepSeek-R128.5
o1 Preview28.5
Claude 3.5 Sonnet (June 2024)26.0
Claude 3.5 Sonnet (Oct 2024)25.3
GPT-4o (source unspecified)25.0
DeepSeek-V323.7
GPT-4 Turbo22.9
o1-mini22.2
Gemini 1.5 Pro (source unspecified)21.9
Opus 321.5
DeepSeek-Coder-V2-Instruct21.2
Llama 3.1 405B Instruct19.8
Claude 3 Sonnet17.0
Llama 3.1 70B Instruct17.0
Qwen2 72B Instruct17.0
Mixtral 8x22B Instruct16.3
Llama 3 70B Instruct14.6
Loading Atlas data…