nisaefendioglu/synthetic-sensitive-data-in-source-code-n300

Synthetic Sensitive Data in Source Code (N=300)

Synthetic dataset of 300 source-code / config snippets containing hardcoded secrets and PII. Every sample includes at least one sensitive finding (no clean negatives).

Designed for evaluating local masking, secret detection, and OWASP LLM02 — Sensitive Information Disclosure scenarios in AI-assisted coding workflows.

Version 1.2: multi_secret (and related) samples label every secret present in code_text (complete ground truth).

All values are synthetic / fake. Do not treat them as real credentials.

Files

Categories (N=300)

Languages include Python, JavaScript, Java, C#, Kotlin, Go, Bash, env, YAML, and JSON.

Schema

• —id — sample id (SDS-####)
• —category — primary category
• —language — snippet language
• —code_text — raw code/config (model input)
• —sensitive_count — number of labeled secrets
• —finding_types — secret types joined by | (CSV)
• —sensitive_findings — ground-truth list (JSON only)

OWASP alignment is dataset-level (LLM02); there is no per-row OWASP column.

Intended use

• —Secret / PII detection benchmarks
• —Local masking and reverse-masking evaluation
• —Prompt/code leakage experiments with LLMs

Notes

• —Fully synthetic; reproducible with seed=42
• —Average secrets per sample ≈ 1.49
• —Not a production vulnerability corpus

Citation

If you use this dataset, please cite the related thesis / paper work by Nisa Nur Efendioğlu.