Datasetpublic

Muneerali199/RakshakAI-v2-Phase-B

RakshakAI v2 - Phase B Dataset Description 192K balanced multi-task security SFT dataset (50/50 vulnerable/clean split). Built from NVD, OSV, BigVul, PrimeVul, Devign, and OWASP Benchmark. Structure pack/ - Axolotl-packed format (with _text field), ready for QLoRA training instruct/ - Chat format with system/user/assistant meta/ - Raw SecuritySample records (programming field preserved) benchmark_hard/ - 500-sample hard benchmark (13 CWE classes… See the full description on the dataset page: https://huggingface.co/datasets/Muneerali199/RakshakAI-v2-Phase-B.

sourceHugging Faceupdated 3mo agoView on Hugging Face

0likes38downloads

Dataset Card

RakshakAI v2 - Phase B Dataset

Description

192K balanced multi-task security SFT dataset (50/50 vulnerable/clean split). Built from NVD, OSV, BigVul, PrimeVul, Devign, and OWASP Benchmark.

Structure

—pack/ - Axolotl-packed format (with _text field), ready for QLoRA training
—instruct/ - Chat format with system/user/assistant
—meta/ - Raw SecuritySample records (programming field preserved)
—benchmark_hard/ - 500-sample hard benchmark (13 CWE classes, unseen repos)

Format

Each sample: system/user/assistant with 5-step CoT reasoning trace + JSON response.

Languages

C (100K), text/CVE (55K), JS (9K), Java (8K), PHP (6K), Python (4K), Rust (3K), Go (2.5K), C# (2K), Swift (1.5K), Ruby (1K), and more.

Data Sources (Not in Benchmark)

All code sourced from public security datasets. Benchmark samples hand-crafted from 2025-2026 CVEs in repos NOT in training data.