CWE-bench v2: Expanded Coverage, Improved Quality, and Greater Transparency
We are releasing CWE-bench v2, a major update to the CWE-bench dataset hosted on Hugging Face under collinear-ai/cwe-bench. The v2 release reflects our ongoing effort to make benchmark data for Common Weakness Enumeration (CWE) research more complete, more reliable, and easier to audit.
The updated dataset, last refreshed on November 3, 2026, contains 100 entries spanning 75 unique CWEs, with contributors from 5 sources. This release focuses on three priorities: more coverage, higher quality, and greater transparency.
What's New in v2
Expanded Coverage
CWE-bench v2 broadens the range of CWEs and source material represented in the benchmark. The dataset now covers 75 distinct weakness types, giving researchers a wider base for evaluating detection and classification systems across the CWE taxonomy.
Improved Data Quality
The v2 release includes refinements to labels, metadata, and record consistency, aimed at reducing noise in downstream evaluation. These improvements help ensure that benchmark results reflect genuine model capability rather than data artifacts.
Greater Transparency
We have made it easier to understand how each entry was sourced and structured, so users can audit and reproduce their evaluations. The Hugging Face dataset page exposes update timestamps and composition details for quick verification.
Dataset at a Glance
| Attribute | Value |
| --- | --- |
| Dataset | collinear-ai/cwe-bench |
| Version | v2 |
| Last updated | 2026-11-03 09:21:27 GMT |
| Total entries | 100 |
| Unique CWEs | 75 |
| Contributors / sources | 5 |
Why It Matters in 2026
As LLM-based security tooling matures, the demand for high-quality, well-documented benchmarks has grown. CWE-bench v2 aims to serve as a dependable reference for teams building and evaluating vulnerability detection systems, with transparency practices that support reproducible research.
Access the Dataset
The dataset is available on Hugging Face: collinear-ai/cwe-bench. We welcome feedback and contributions from the community as we continue to grow the benchmark in future releases.