File size: 3,437 Bytes
4535fe4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 | ---
pretty_name: SciCode-Verified
license: apache-2.0
language:
- en
annotations_creators:
- expert-generated
source_datasets:
- extended
task_categories:
- text-generation
size_categories:
- n<1K
tags:
- code
- science
- benchmark
- code-generation
- scientific-coding
- llm-evaluation
configs:
- config_name: default
data_files:
- split: test
path: data/problems_test.jsonl
---
# SciCode-Verified
SciCode-Verified is the corrected, human-verified release of the
[SciCode](https://github.com/scicode-bench/SciCode) scientific-code-generation benchmark.
A problem-by-problem audit identified 263 defects in the 65-problem SciCode test split and
corrected every confirmable defect. The released evaluation set contains 64 main problems and
287 scored subproblems; one original problem is excluded because its specification does not
determine a unique, verifiable answer.
- **Paper:** [SciCode-Verified: How Benchmark Defects Underestimated the Scientific-Coding Ability of Language Models](https://arxiv.org/abs/2608.04975)
- **Code, evaluation harness, and audit trail:** [github.com/flyingwagner/scicode-verified](https://github.com/flyingwagner/scicode-verified)
- **Original benchmark:** [SciCode](https://arxiv.org/abs/2407.13168)
## Files
| File | Purpose |
|---|---|
| `data/problems_test.jsonl` | Corrected prompts and problem specifications, one main problem per line |
| `test_data_cleaned.h5` | Corrected frozen grading targets used by the evaluation harness |
| `manifest.json` | Release version, problem order, and MD5 checksums |
| `LICENSE` | Apache License 2.0 inherited from the upstream benchmark |
Release `v2` checksums:
```text
5c604d8dbf52642bd94e13b92c8f52eb data/problems_test.jsonl
2b41a7df40ddc23ce651ec05b8ecb6f8 test_data_cleaned.h5
```
## Load the problem specifications
```python
from datasets import load_dataset
dataset = load_dataset("shhu2001/SciCode-Verified", split="test")
print(dataset[0])
```
## Download the grading targets
```python
from huggingface_hub import hf_hub_download
h5_path = hf_hub_download(
repo_id="shhu2001/SciCode-Verified",
repo_type="dataset",
filename="test_data_cleaned.h5",
)
```
The full generation, grading, multi-environment evaluation, and integrity-check workflow is
documented in the [GitHub repository](https://github.com/flyingwagner/scicode-verified). Model
outputs and evaluation logs are intentionally not included in this dataset repository.
## License and attribution
SciCode-Verified is derived from SciCode and redistributed under the Apache License 2.0. Please
cite both the original SciCode benchmark and SciCode-Verified.
```bibtex
@article{hu2026scicodeverified,
title = {{SciCode-Verified}: How Benchmark Defects Underestimated the Scientific-Coding Ability of Language Models},
author = {Hu, Sihan and Huang, Lyuhan and Deng, Youjin and Chen, Kun},
year = {2026},
eprint = {2608.04975},
archivePrefix = {arXiv},
primaryClass = {cs.SE},
url = {https://arxiv.org/abs/2608.04975}
}
@article{tian2024scicode,
title = {{SciCode}: A Research Coding Benchmark Curated by Scientists},
author = {Tian, Minyang and Gao, Luyu and Zhang, Shizhuo Dylan and Chen, Xinan and others},
year = {2024},
eprint = {2407.13168},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2407.13168}
}
```
|