Cve Train V1.1: 18,783 Agent Traces for CVE Reproduction

18,783 distilled agent traces for CVE reproduction tasks, generated using the Claude Opus 4.5 model with a Mini SWE-Agent harness through the CVE-Factory pipeline. This is an expanded version of the cve_train dataset, adding approximately 3,000 tasks from the cve_tasks_3k_compressed set. The dataset was authored by Luoberta and last updated on March 27, -2026.

Use Cases

• Training or fine-tuning AI agents for automated CVE reproduction based on the distilled agent traces.
• Benchmarking agent performance on software security tasks based on the described CVE task coverage.
• Analyzing patterns in AI-assisted vulnerability exploitation workflows based on the synthetic agent traces.
• Researching the effectiveness of large language models in security contexts based on the Claude Opus generation method.

Strengths

• Contains 18,783 samples, a significant increase from the previous version's ~4,000.
• Traces were generated using a specified methodology involving Claude Opus 4.5 and a Mini SWE-Agent harness.
• Includes an additional ~3,000 tasks from a named source (cve_tasks_3k_compressed).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect temporal or source bias inherent to the synthetic generation method.

Provenance

Source

huggingface

Collection Method

Synthetically generated using Claude Opus 4.5 with a Mini SWE-Agent harness through the CVE-Factory pipeline.

Time Range

null

Freshness

Last updated 2026-03 27 07:11:43; freshness should be verified.

Geography

null