CVE Train: 4,078 Agent Traces for Security Vulnerability Reproduction

CVE-Factory Agent Traces is a dataset of 4,078 distilled agent traces generated by Luoberta for 887 CVE reproduction tasks. The traces were created using Claude Opus 4.5 with a Mini SWE-Agent harness. The dataset was last updated on February 4, 2026.

Use Cases

• Fine-tuning language models for security agent performance based on the described agent traces.
• Benchmarking AI agents on CVE reproduction tasks using the provided distilled traces.
• Studying synthetic agent behavior in software engineering security contexts.
• Training models for automated vulnerability analysis and patching workflows.

Strengths

• Contains 4,078 distilled agent traces for a substantial scope of tasks.
• Covers 887 distinct CVE reproduction tasks, providing a focused security domain corpus.
• Traces were generated using a specified model (Claude Opus 4.5) and harness (Mini SWE-Agent), indicating a controlled generation method.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but other scale details like file formats and size are unknown.
• The dataset excludes the simpler 'trainset-2' split, which may limit task diversity.

Provenance

Source

Luoberta on Hugging Face, part of the CVE-Factory pipeline.

Collection Method

Generated using Claude Opus 4.5 with a Mini SWE-Agent harness.

Time Range

null

Freshness

Last updated 2026-02 04:32:15; freshness should be verified.

Geography

null