SciCodePile: leaderboard
The executable benchmark of SciCodePile (Singapore Management University and others, 2026): 200 HumanEval-style tasks derived from pure Python functions in scientific code repositories, each run in a sandbox with stubbed dependencies against a synthesized and manually reviewed test harness. The score is Pass@1 in percent; higher is better.
Metric: Pass@1 (%) on the 200 executable scientific Python function-generation tasks, each run in a sandbox with stubbed dependencies against its test harness (HumanEval protocol); higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 Mini | 12.3 | #389 |
| 2 | O3 Mini | 10.5 | #533 |
| 3 | DeepSeek R1 | 8.5 | #473 |
| 4 | Qwen 3 14B (Reasoning) | 8.3 | #1233 (Qwen 3 14B) |
| 5 | GPT-4o | 7.5 | #741 |
| 6 | Qwen 2.5 Coder 7B Instruct | 6.1 | #3852 |
| 7 | GPT-5 | 5.7 | #167 |
| 8 | DeepSeek R1 Distill Qwen 1.5B | 2.7 | #9195 |
| 9 | starcoder2-3B | 0.5 | #9791 |
| 10 | starcoder2-7B | 0.3 | #9496 |
| 11 | starcoder2-15B | 0.3 | #8482 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=scicodepile · How It Works · Data refreshed daily, snapshot 2026-11-13.