Thinking-optimized web-security SFT: 49% <think>, recitation dropped, grounded+authored, leakage-safe

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+train.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,40 @@
+---
+language: [en]
+license: gpl-2.0
+task_categories: [text-generation]
+tags: [cybersecurity, application-security, bug-bounty, pentest, thinking, reasoning, tool-calling, jsonl, qwen]
+size_categories: [1K<n<10K]
+---
+
+# OffsecCyber-Thinking — reasoning-optimized web-security SFT
+
+A **thinking-mode** supervised fine-tuning set for web application security / bug-bounty models (Qwen3-style `<think>` + tool-calling). Reasoning-dominant and grounded: CVE-metadata **recitation was deliberately dropped** to reduce hallucination.
+
+## Composition (9,417 records)
+| grounding | count | source |
+|---|---:|---|
+| grounded | 4,001 | Exploit-DB verified webapps, PoC extracted from the real exploit files |
+| authored_curated | 2,500 | offensive/web-appsec instruction data |
+| real_report | 2,000 | HackerOne-style reports |
+| authored | 643 | synthetic tail-class + modern-stack + negatives (think + tool-calling) |
+| grounded_lab | 273 | PortSwigger labs |
+
+- **`<think>` share: 49%** · **negatives: 7.5%** · tool-calling trajectories included.
+- Splits: train 8,475 / valid 470 / test 472, **leakage-safe** (grouped by CVE/EDB/report-id/hash), 0 cross-split leaks, 0 fabricated/forward CVEs, near-dup ≤3/template.
+- Each record tags `metadata.grounding / license / provenance / has_think / is_negative / cvss_source / tool_output`.
+
+## ⚠️ Use in production (read before training/deploying)
+This set trains **reasoning and judgment**, not a fact oracle. To avoid hallucination:
+1. **RAG for facts** — CVSS is estimated, not authoritative; CVE recall is unreliable at ≤8B. Supply CVE/CVSS/version truth from NVD/GHSA at inference.
+2. **Real tools, never invented output** — `metadata.tool_output="simulated"` marks learned tool results; execute tools for real in production.
+3. **Human-in-the-loop on exploitation** — PoCs are drafts to verify in a sandbox.
+
+## License / provenance (IMPORTANT)
+Mixed provenance, tagged per record in `metadata.license`:
+- **`GPL-2.0-or-later`** (4,001, provenance=exploit-db) — the card license reflects this (most restrictive).
+- **`review-upstream`** (4,773, h1/portswigger/web_offensive) — confirm each upstream source's terms before redistribution.
+- **`synthetic-authored`** (643) — project-authored.
+
+For a **closed-source** product: training is generally permissible, but redistribution of GPL-derived data/weights can carry obligations — get legal review, or train a shippable model on the non-GPL (`provenance != exploit-db`) subset.
+
+🤖 Generated with [Claude Code](https://claude.com/claude-code)

manifest.json

@@ -0,0 +1,25 @@
+{
+  "name": "Security SFT \u2014 THINKING-optimized (reasoning-dominant, recitation dropped)",
+  "seed": 240517,
+  "records": 9417,
+  "splits": {
+    "train": 8475,
+    "valid": 470,
+    "test": 472
+  },
+  "think_share": 0.493,
+  "negative_share": 0.075,
+  "grounding_mix": {
+    "grounded": 4001,
+    "authored_curated": 2500,
+    "real_report": 2000,
+    "authored": 643,
+    "grounded_lab": 273
+  },
+  "license_mix": {
+    "review-upstream": 4773,
+    "GPL-2.0-or-later": 4001,
+    "synthetic-authored": 643
+  },
+  "notes": "recitation (qwen_cve/NVD) dropped; all <think> records kept; curated non-think capped. RAG for CVE/CVSS facts; real tools only; human-verify PoCs."
+}

train.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04628b44f2d699936e55423bbdc96daf55b470f1074ed3ed86b1949259442556
+size 30434383