ArkhAngelLifeJiggy/cve_train

CVE-Factory Agent Traces

This dataset contains 4,078 distilled agent traces from 887 CVE reproduction tasks (the `trainset/` split), generated using Claude Opus 4.5 with a Mini SWE-Agent harness. See CVE-Factory for the full pipeline.

Note: CVE-Factory also provides a trainset-2/ split with additional simpler tasks, which is not included in this training dataset.

🚀 Training Results

Fine-tuning on this dataset yields dramatic improvements across security benchmarks:

Key findings:

• —~6.8× improvement on LiveCVEBench (5.29% → 35.79%)
• —~4.2× improvement on PatchEval (5.66% → 23.58%)
• —~2.3× improvement on Terminal-Bench (12.50% → 28.75%)
• —[Abacus-cve](https://huggingface.co/Luoberta/Abacus-cve) (32B) outperforms Qwen3-Coder-480B, MiniMax-M2, and Claude Sonnet 4
• —Approaches Claude Sonnet 4.5 level on security tasks

Dataset Format

Each line is a JSON object with:

{
  "task_id": "cve-2017-15197.2-of-5.2026-01-25__22-10-14",
  "is_resolved": true,
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."},
    ...
  ]
}

• —task_id: Unique task identifier (CVE ID + trace index + timestamp)
• —is_resolved: Whether the task was successfully completed
• —messages: Conversation history in standard chat format (system/user/assistant turns)

Usage

from datasets import load_dataset

dataset = load_dataset("Luoberta/cve_train")

# Access a sample
sample = dataset["train"][0]
print(f"Task: {sample['task_id']}")
print(f"Resolved: {sample['is_resolved']}")
print(f"Turns: {len(sample['messages'])}")

Related Resources

• —[Abacus-cve](https://huggingface.co/Luoberta/Abacus-cve) - Model fine-tuned on this dataset
• —[Leaderboard](https://livecvebench.github.io/) - Live rankings on LiveCVEBench
• —[LiveCVEBench](https://github.com/livecvebench/LiveCVEBench-Preview) - Security vulnerability benchmark
• —[CVE-Factory](https://github.com/livecvebench/CVE-Factory) - The multi-agent system that generated these traces

Citation

@misc{luo2026cvefactory,
  title={CVE-Factory: Scaling Expert-Level Agentic Tasks for Code Security Vulnerability}, 
  author={Xianzhen Luo and Jingyuan Zhang and Shiqi Zhou and Rain Huang and Chuan Xiao and Qingfu Zhu and Zhiyuan Ma and Xing Yue and Yang Yue and Wencong Zeng and Wanxiang Che},
  year={2026},
  eprint={2602.03012},
  archivePrefix={arXiv},
  primaryClass={cs.CR},
  url={https://arxiv.org/abs/2602.03012}
}