Denash/cevud-pipeline-dataset

CEVuD Pipeline Dataset (VUDENC-based)

HuggingFace-ready dataset card for the VUDENC-based pipeline corpus. Published at Denash/cevud-pipeline-dataset.

This dataset is the held-out corpus for the CEVuD gate study. It tunes and evaluates the full three-stage pipeline (Semgrep + small model + gating + LLM) on code the classifier never trained on. It is derived from VUDENC (Wartschinski et al., Information & Software Technology, 2022).

The CEVuD Stage-2 classifier trained on this dataset is Denash/codebert-vuln-classifier.

---

Dataset Summary

Property	Value
HF Dataset ID	Denash/cevud-pipeline-dataset
Source dataset	DetectVul/Vudenc
Original paper	Wartschinski et al., "Vulnerability Detection with Deep Learning on a Natural Codebase", Information & Software Technology, 2022
Language	Python
Task	Binary sequence classification (vulnerable vs. safe) at function level
Label scheme	0 = safe, 1 = vulnerable (derived from per-line annotations)
Total samples	7,680
Projects (vulnerability types)	12
Vulnerable samples	1,587 (20.7%)
Safe samples	6,093 (79.3%)
License	Same as VUDENC (check upstream)

---

Dataset Creation

Source Data

VUDENC is a HuggingFace dataset containing Python functions with per-line vulnerability annotations. Each row contains:

• raw_lines: original source code lines
• lines: tokenized/normalized code lines
• label: per-line binary vulnerability label (1/0)
• type: per-line vulnerability category string

The seven vulnerability categories in VUDENC are:

1. SQL injection (sql)
2. Cross-Site Scripting (xss)
3. Command injection (command)
4. Cross-Site Request Forgery (xsrf)
5. Remote Code Execution (remote_code_execution)
6. Path Disclosure (path_disclosure)
7. Open Redirect (open_redirect)

Curation Pipeline

The raw VUDENC data is converted through a pipeline implemented in src/scripts/convert_vudenc.py.

Stage 1 — Function-level labeling

1. A function is labeled vulnerable (1) if any of its lines is labeled vulnerable. Otherwise safe (0).
2. The dominant vulnerability type is extracted from the most-common per-line type string.

Stage 2 — Source reconstruction

3. raw_lines are joined back into a full function string.

Stage 3 — Minimum signal filter

4. Functions with fewer than 2 lines of real code signal are dropped.

Stage 4 — Deduplication

5. Exact deduplication of function strings.
6. Contradiction removal: drop identical text appearing with both labels.

Stage 5 — Project grouping

7. Samples are grouped by their dominant vulnerability type (e.g. vudenc_sql, vudenc_xss) to form logical "projects" for leakage-safe splitting.

Stage 6 — Optional benign controls

8. Verified-benign control samples can be merged in as local_source projects for the gate study, enabling precision/recall computation.

Stage 7 — Splitting

9. Project-level splitting with seed=42. No vulnerability type appears in more than one split.

---

Dataset Structure

Field	Type	Description
sample_id	str	Unique identifier (e.g. vudenc::sql::000001::abc123)
project	str	Vulnerability type group (e.g. vudenc_sql)
source_code	str	Full function source code
label	int	0 = safe, 1 = vulnerable
vulnerability_type	str	Dominant vulnerability type (e.g. sql)
file_path	str	Synthetic path (inline_snippet.py)
function_name	str	Inferred function name
start_line	int	Function start line (1-based)
end_line	int	Function end line (1-based)
sample_subtype	str	vulnerable or benign

Note: VUDENC ships no repository or commit metadata, so provenance fields (repo_url, commit_id, target_commit, cve_id, cvss_score, diff_with_context, fixed_code) are present for schema consistency but are empty.

---

Data Splits

Split	Samples	Vulnerable	Safe	Projects
Train	1,863	456	1,407	8
Validation	4,996	1,026	3,970	2
Test	821	105	716	3

---

Per-Project Breakdown

Project	Total	Vulnerable	Safe
vudenc_AnnAssign'	2	0	2
vudenc_Assert'	22	5	17
vudenc_Assign'	2,952	850	2,102
vudenc_AsyncFunctionDef'	18	1	17
vudenc_AugAssign'	19	3	16
vudenc_Condition	789	101	688
vudenc_Expr'	1,255	295	960
vudenc_For	32	4	28
vudenc_FunctionDef'	2,044	176	1,868
vudenc_Import'	85	25	60
vudenc_ImportFrom'	197	98	99
vudenc_Return'	265	29	236

---

Key Statistics

• Class distribution: ~1 : 4 vulnerable/safe overall
• Vulnerability type diversity: 7 categories
• Most common type: vudenc_Assign (2,952 samples, 850 vulnerable)
• Least common type: vudenc_AnnAssign' (2 samples, 0 vulnerable)
• Function size: Average ~20–40 lines per function

---

Intended Use

This dataset is intended for:

• Gate tuning: Selecting the linear gate weights (W₁, W₂, T_escalation) that maximize F2 on the validation split.
• Pipeline evaluation: Measuring the full CEVuD pipeline's recall, precision, TRR, and cost reduction on unseen data.
• Benchmarking: Comparing CEVuD against baseline strategies (Semgrep only, small model only, logistic regression, etc.).

It is not intended for:

• Training the small model (use the CVEfixes-based training dataset instead)
• Standalone vulnerability detection
• Training on the test split (reserved for final evaluation)

---

Limitations

• Function-level granularity: A function is labeled vulnerable if any line is vulnerable. This may over-label functions containing both vulnerable and safe code.
• Python-only: All samples are Python functions.
• Limited vulnerability diversity: Only 7 vulnerability types, with vudenc_Assign dominating (2,952 of 7,680 samples).
• No repository context: VUDENC provides isolated functions, not full repositories. Cross-file dependencies are not captured.
• Class imbalance: The safe class dominates (79.3%), affecting precision/recall trade-offs.
• Temporal bias: Samples were collected at a specific point in time and may not reflect current vulnerability patterns.
• Label noise: Per-line annotations may contain errors.

---

Citation

@misc{cevud2026,
  title={CEVuD: Cost-Effective Vulnerability Detection via Gated Static-Neural Reasoning},
  author={CEVuD Authors},
  year={2026},
  note={Dataset: Denash/cevud-pipeline-dataset; Model: Denash/codebert-vuln-classifier}
}

@article{wartschinski2022vudenc,
  title={Vulnerability Detection with Deep Learning on a Natural Codebase},
  author={Wartschinski, Laura and others},
  journal={Information and Software Technology},
  year={2022}
}

---

Related Resources

Resource	Link
Model	Denash/codebert-vuln-classifier
Training dataset (CVEfixes)	Denash/cevud-training-dataset
Source dataset (VUDENC)	DetectVul/Vudenc
CEVuD GitHub	https://github.com/Denash/CEVuD

Point of contact: Open an issue on the CEVuD GitHub repository.