Datasetpublic

nyxspecter4/kin-cyber-dpo-v2

KIN Cybersecurity DPO v2 Preference Dataset Empirically mined and zero-leak sanitized preference dataset for training cybersecurity and agentic code repair models. Dataset Summary Total DPO Pairs: 1,635 (Updated 2026-09-07) Baseline v4 pairs: 1,495 Hermetic expansion (v5): +140 pairs (AST-invariant vulnerability repair, CWE-79 XSS guards, CWE-89 SQLi, CWE-22 Path Traversal, and supply chain integrity) Format: Direct Preference Optimization (DPO) schema: {prompt… See the full description on the dataset page: https://huggingface.co/datasets/nyxspecter4/kin-cyber-dpo-v2.

sourceHugging Faceupdated 14d agoView on Hugging Face

0likes519downloads

Dataset Card

KIN Cybersecurity DPO v2 Preference Dataset

Empirically mined and zero-leak sanitized preference dataset for training cybersecurity and agentic code repair models.

Dataset Summary

Total DPO Pairs: 1,635 (Updated 2026-09-07) Baseline v4 pairs: 1,495 Hermetic expansion (v5): +140 pairs (AST-invariant vulnerability repair, CWE-79 XSS guards, CWE-89 SQLi, CWE-22 Path Traversal, and supply chain integrity)

Format: Direct Preference Optimization (DPO) schema: {prompt, chosen, rejected, metadata}

Schema

Column	Type	Description
prompt	string	The input scenario (code diff, CVE description, agent trace)
chosen	string	Verified brief with working check / PoC
rejected	string	Vague filing without verification
metadata	json	CWE, target, vector, date

Files

File	Size	Description
`dpo.jsonl`	1.5 MB	Full dataset — 1,635 DPO pairs (canonical)
`dpo-clean.jsonl`	476 KB	Cleaned/deduplicated subset (431 pairs)
`sft.jsonl`	943 KB	Supervised fine-tuning format
`train.jsonl`	472 KB	Training split
`grpo_prompts.jsonl`	2.3 KB	GRPO prompt seeds

Note: The data_files config above points HF Dataset Viewer to dpo.jsonl (the full 1,635-pair dataset). Previously, HF auto-detected dpo-clean.jsonl (alphabetically first), showing only 431 rows.

Provenance

—Mined from: CVE databases, monk-bounty evidence envelopes, code repair traces
—Sanitization: zero-leak filtering, dedup by prompt hash
—Related model: nyxspecter4/kin-cybersecurity-3b
—Related GGUF: nyxspecter4/kin-cybersecurity-3b-gguf
—Demo Space: nyxspecter4/kin-cybersec

DPO Contrast Examples

Example 1: SQL Injection (CWE-89)

Prompt: A PR diff adding cursor.execute(f"SELECT * FROM users WHERE id = {user_id}")

Chosen: "CLAIM: SQL injection via f-string interpolation. BLAST: Full DB read/write, auth bypass. WHY: CWE-89, user input reaches execute() without parameterization. VERIFY: curl 'http://target/api?id=1 OR 1=1' returns all users. CONFIDENCE: High — falsified if parameterized query is used."

Rejected: "This code looks dangerous and might have a security issue with the database query."

License

MIT