openSUSE/cve-backport-codegen-dataset

CVE Backport Code Generation Dataset

Per-hunk code generation dataset for CVE security patch backporting, derived from openSUSE Build Service maintenance patches.

Task

Given a region of vulnerable source code and a description of the upstream CVE fix, the model outputs the fixed version of the code. A programmatic diff then produces the final patch. This plays to LLM strengths in code completion and avoids format-sensitivity issues with direct diff generation.

Starting with v4, some examples also include a second turn where the model generates a test case that triggers the vulnerability in the original code and passes after the fix.

Files

File	Examples	Description
train.jsonl	36,168	v4 (36,166 examples, 770 with test cases)
eval.jsonl	1,834	Held-out evaluation set (unchanged across versions)

Version History

Version	Examples	Packages	What changed
v4	36,168	145	+501 examples from source matching fix, 772 five-turn examples with CVE test cases, 99.8% SPDX license coverage
v3	35,778	~120	Cleaned v2+augmentation: removed 186 toxic examples (XML test data, DH prime tables, assembly blobs that crashed training at epoch 0.46)
v2	24,452	91	Synthetic cross-version augmentation (+7,445 adapted-tier examples), more packages mined from OBS
v1	17,007	91	Initial dataset: per-hunk code generation from openSUSE maintenance patches

Format

3-turn examples (codegen only, 35,396 in v4)

{
  "messages": [
    {"role": "system", "content": "You are a security patch backporting assistant..."},
    {"role": "user", "content": "## File: lib/url.c\n## Lines: 100-130\n\n```c\n{vulnerable code}\n```\n\n## Fix\nCVE-2024-1234: description"},
    {"role": "assistant", "content": "{fixed code}"}
  ],
  "metadata": {
    "cve_id": "CVE-2024-1234",
    "package": "curl",
    "language": "c",
    "tier": "identical",
    "suse_license": "MIT"
  }
}

5-turn examples (codegen + test, 772 in v4)

{
  "messages": [
    {"role": "system", "content": "You are a security patch backporting assistant..."},
    {"role": "user", "content": "## File: lib/url.c\n## Lines: 100-130\n..."},
    {"role": "assistant", "content": "{fixed code}"},
    {"role": "user", "content": "Write a test case that:\n1. Triggers the vulnerability...\n2. Passes after applying your fix"},
    {"role": "assistant", "content": "{test case code}"}
  ],
  "metadata": {"...", "has_test": true}
}

Coverage

• 2,300+ CVEs across 145 packages
• Languages: C (80%), Python (7%), C++ (3%), Shell, Perl, JavaScript, Java, Go, PHP, Ruby
• Tiers: identical (patch applies directly), adapted (context differs), synthetic-adapted (cross-version augmentation)
• 99.8% per-example SPDX license metadata

Licenses

All training examples are derived from publicly available openSUSE Build Service maintenance patches. Per-example license metadata is in metadata.suse_license. The most common licenses are:

LGPL-2.1-or-later, OpenSSL, GPL-2.0-only, MIT, PHP-3.01, Apache-2.0, GPL-2.0-or-later, BSD-3-Clause, GPL-3.0-or-later, BSD-2-Clause, Python-2.0, MPL-2.0, ImageMagick

Model

Trained models are at openSUSE/CVE-Backport-Qwen2.5-Coder-32B.

The CLI tool for using the model is at github.com/openSUSE/cve-backport-tool.

Reproducibility

Each versioned training file corresponds to a specific model version. To reproduce a model:

1. Download the matching train-vN.jsonl and eval.jsonl
2. Use the Teapot composable training pipeline:

   teapot compose configs/cve-backport.config
   teapot train configs/cve-backport.config --backend qlora-hf
   teapot eval configs/cve-backport.config