True2456/cybersecurity-theory-sft-gemma12b

Cybersecurity Theory SFT (Gemma 12B pack)

Curated 21,265-row cybersecurity theory instruction pack for LoRA supervised fine-tuning. Each example is a single-turn user → assistant pair covering offensive/defensive concepts, frameworks, CTF reasoning, vulnerability catalogs, and security tooling literacy — without agent tool traces or multi-turn harness data.

Paired MLX LoRA adapter trained on this pack (Nemotron 3 Super export):
True2456/nemotron-3-super-120b-cybersecurity-theory-lora-mlx

Rundown

Item	Value
Purpose	Stage-1 theory specialist for the Mati MultiLoRA stack — domain grounding before agentic / code-as-action stages
Rows (full pack)	21,265
Train split	~19,138 (90%, stratified by source, seed 42)
Valid split	~2,127 (10%, same stratification)
Max completion length	6,000 characters (paragraph-aware truncation during curation)
Suggested base models	mlx-community/gemma-4-12b-it-bf16, mlx-community/Nemotron-3-Super-120B-A12B-MLX-6bit
Build script	scripts/build_theory_packs.py
Upstream download	scripts/download_tier1.sh

Dataset lineage

Tier-1 upstream (download via scripts/download_tier1.sh)
  RISys-Lab/RedSage-Seed
  trendmicro-ailab/Primus-Seed | Primus-Instruct | Primus-Reasoning
  justinwangx/CTFtime (+ CTFtime-unrolled)
  CISA KEV JSON (via theory/substitutes-for-cybersecurity-1m/)
        │
        ▼  scripts/build_theory_packs.py
curated/theory_gemma12b/
  theory_gemma12b_messages.jsonl   (chat messages)
  theory_gemma12b_train.jsonl      (alpaca instruction/input/output)
  manifest.json
        │
        ▼  90/10 stratified split (seed=42, by source)
splits/
  train_messages.jsonl / valid_messages.jsonl
  train_alpaca.jsonl   / valid_alpaca.jsonl

Composition (by source)

Source	Rows	Role
RedSage-Seed	14,846	Frameworks, general cyber knowledge, offensive skills, Kali tooling, filtered CLI references
Primus Reasoning (CTIBench / DeepSeek-R1 style)	2,489	Longer cybersecurity reasoning traces
CTFtime writeups	2,345	Competition writeups reframed as teaching explanations
CISA KEV	800	Known Exploited Vulnerabilities catalog Q&A
Primus Instruct	785	Alert explanation, cmd analysis, Terraform misconfig, SIEM-style queries, security doc QA
Total	21,265

RedSage subset breakdown (within 14,846): general 6,924; framework 3,715; skills 2,490; Kali 1,023; CLI 694.

Primus Instruct subsets: general 306; cmd_analysis 100; alert_explanation 100; terraform_misconfiguration_scan 96; security_doc_qa 92; security_event_query_generation 91.

Row schema

Messages format (splits/train_messages.jsonl, splits/valid_messages.jsonl):

{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "source": "redsage",
  "subset": "framework",
  "fingerprint": "abc123..."
}

Alpaca format (splits/train_alpaca.jsonl, splits/valid_alpaca.jsonl):

{
  "instruction": "...",
  "input": "",
  "output": "...",
  "source": "redsage",
  "subset": "framework",
  "fingerprint": "abc123..."
}

Rebuilding from source

1. Download approved Tier-1 datasets (requires hf auth login):

   bash scripts/download_tier1.sh
   

2. Build the pack locally:

   python3 scripts/build_theory_packs.py
   

3. See docs/ACCESS.md for download status and docs/SUBSTITUTES.md for public replacements when CyberSecurity-1M is unavailable.

Note: This Hugging Face release ships the curated SFT pack only, not the multi-gigabyte upstream parquet dumps. Rebuilders must fetch upstream datasets under their respective licenses.

Research foundations

This pack follows the cybersecurity LLM training literature and publicly released corpora below. Mati_Train curates and reformats upstream data into instruction-tuning rows; it does not redistribute raw upstream dumps.

Work	Citation	Used for
PRIMUS	Yang et al., EMNLP 2025 — arXiv:2502.11191	Primus-Instruct, Primus-Reasoning subsets
RedSage	Suryanto et al., ICLR 2026 — arXiv:2601.22159	RedSage-Seed knowledge / skills / tools taxonomies
LoRA	Hu et al., 2021 — arXiv:2106.09685	Parameter-efficient fine-tuning methodology
CTFtime	justinwangx/CTFtime	CTF writeup teaching rows
CISA KEV	CISA Known Exploited Vulnerabilities catalog	Vulnerability literacy Q&A

Recommended use

• LoRA / QLoRA SFT of cybersecurity theory specialists (explanation, mentoring, concept walkthroughs)
• Held-out valid_* splits for mlx-lm eval during training
• Domain expert slot in multi-adapter / MoE routing stacks

Out of scope

• Not an agentic tool-calling dataset (no bash / file / patch trajectories)
• Not a substitute for authorized penetration testing or production security operations
• Not guaranteed factually current — CVE/KEV entries and tooling evolve rapidly

Disclaimers

Educational / Research / Personal Use Only. See LICENSE in this repository.

1. Permitted use: classroom instruction, academic research, personal learning, and private non-commercial experimentation.
2. Prohibited use: commercial deployment, paid services, unauthorized offensive security operations, or attacks against systems you do not own or lack explicit written permission to test.
3. Offensive content: rows include offensive methodology, exploit primitives, and CTF techniques for defensive education. Possession of knowledge does not authorize illegal activity.
4. Accuracy: content is derived from public sources and automated curation. Verify against authoritative references (vendor advisories, MITRE, CISA) before operational decisions.
5. Upstream licenses: RedSage-Seed, PRIMUS, CTFtime, and CISA data each carry their own terms. Comply with upstream licenses when rebuilding from source.

Files in this release

Path	Description
splits/train_messages.jsonl	Training split, chat messages format
splits/valid_messages.jsonl	Validation split, chat messages format
splits/train_alpaca.jsonl	Training split, alpaca format
splits/valid_alpaca.jsonl	Validation split, alpaca format
manifest.json	Pack metadata, source counts, split info
scripts/build_theory_packs.py	Pack builder (reproducibility)
scripts/download_tier1.sh	Upstream dataset downloader
docs/ACCESS.md	Local access status notes
docs/SUBSTITUTES.md	CyberSecurity-1M substitute mapping
LICENSE	Educational / Research / Personal Use Only

Citation

@misc{true2456_cybersecurity_theory_sft_gemma12b,
  author       = {True2456},
  title        = {Cybersecurity Theory SFT (Gemma 12B pack)},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  note         = {Built from RedSage-Seed, PRIMUS, CTFtime, CISA KEV; see README for papers}
}

License

Released under Educational / Research / Personal Use Only (see LICENSE). Upstream datasets retain their own licenses.