nuhmanpk/rook-secops-decisions

🏰 Rook SecOps Decisions

Typed, calibrated security decisions for training small, fast "System One" decision models. Each case is a piece of security evidence, such as a CVE description, a raw SOC log line, an ATT&CK behaviour, an email, a URL or an LLM prompt. Each case carries one or more typed questions (choice, score, noul) and a gold probability distribution for each question.

The format follows LocalLLaMA/typed-decisions (TypeSafe Jev / Laya System One primitives), so it trains directly with Laya's Kaggle 2×T4 fine-tuning notebook. It is built to train Rook, a security decision model that answers in a single forward pass.

Cases	36,550 (27,566 train / 3,846 validation / 5,138 test)
Typed questions	41,316 (score: 17,618, choice: 8,989, noul: 14,709)
Workflows	12
Sources	9 open datasets (all credited below)

Workflows

Workflow	Decision	Types	Train	Validation	Test
alert_triage	Raw SOC log event → benign / suspicious / malicious	score	2,450	322	289
attack_technique	Adversary behaviour → ATT&CK tactic; targets cloud/SaaS?	choice, noul	615	38	44
cwe_root_cause	CVE description → root-cause CWE (5 options)	choice	1,180	141	159
guidance_domain	NIST guidance excerpt → owning security domain (8 options)	choice	2,523	1,118	850
kev_ransomware	Known-exploited CVE → used in ransomware campaigns?	noul	1,312	164	249
mitigation_select	Adversary behaviour → MITRE mitigation (5 options)	choice	595	63	83
phishing_message	Email or SMS → phishing / scam?	noul	4,029	462	509
phishing_url	URL → phishing link?	noul	2,386	309	305
prompt_attack	LLM prompt → injection / jailbreak attempt?	noul	1,335	140	366
security_incidents	typed-decisions security workflow (5 questions per case)	choice, noul, score	300	0	100
vuln_exploitation	CVE → ATT&CK tactic, network-exploitable, user interaction, privileges required	choice, noul, score	902	98	180
vuln_severity	Vulnerability description → CVSS severity (low / medium / high / critical)	score	9,939	991	2,004

Format

Column	Description
id	unique id
workflow	one of the workflows above
split	train / validation / test
state	JSON: the evidence the model reads
questions	JSON: {question_id: {type, instructions, criteria}} (a System One request body)
gold	JSON: {question_id: {type, label, probabilities, confidence, noul / score}}
labels	JSON: {question_id: argmax label}, for quick viewing
n_questions	number of questions in the case
source_dataset, source_license	provenance of this row
entity	the CVE / technique / log template / document the row is about (used for leak-free splits)
meta	JSON with source fields (e.g. the raw CVSS score) for analysis. Never feed it to a model.

Question types follow the System One primitives:

• choice: pick one of 2–20 named options (criteria maps label → description).
• score: an ordered rubric, lowest level first (criteria is a list). gold.score is the expected level.
• noul: a yes/no statement (criteria keyed false / true). gold.noul is P(true).

Example (alert_triage):

{
  "state": {
    "stream": "aws_cloud_trail",
    "event_type": "audit_log",
    "raw_log": "{\"eventVersion\":\"1.09\",\"userIdentity\":{\"type\":\"IAMUser\",\"principalId\":\"CRED-0011\",\"arn\":\"arn:aws:iam::100000000002:user/CRED-0012\",\"accountId\":\"100000000002\",\"accessKeyId\":\"CRED-0013\",\"userName\":\"USER-0717\"},\"eventTime\":\"2024-07-31T13:41:41Z\",\"eventSource\":\"logs.amazonaws.com\",\"e ..."
  },
  "questions": {
    "triage": {
      "type": "score",
      "instructions": "How should a SOC analyst triage this security event?",
      "criteria": [
        "benign: routine activity that matches no detection rule",
        "suspicious: matches a detection rule but is not part of a confirmed incident",
        "malicious: part of an analyst-confirmed security incident"
      ]
    }
  },
  "gold": {
    "triage": {
      "type": "score",
      "label": "0",
      "probabilities": {
        "0": 0.9,
        "1": 0.1,
        "2": 0.0
      },
      "score": 0.1,
      "confidence": 0.85
    }
  }
}

Load it

import json
from datasets import load_dataset
ds = load_dataset("nuhmanpk/rook-secops-decisions")
row = ds["train"][0]
state, questions, gold = (json.loads(row[k]) for k in ("state", "questions", "gold"))

Train a Rook model with Laya

In Laya's fine-tuning notebook, replace load_dataset("LocalLLaMA/typed-decisions", "all", split="train") with load_dataset("nuhmanpk/rook-secops-decisions", split="train"). Nothing else changes: state, questions and gold have the same JSON shape. Evaluate on test with the notebook's metric cell. The full train split is about 27,566 cases. On 2×T4, start with 1–2 epochs, or subsample by workflow.

How it was built

Label policy. Gold distributions are soft. A label-smoothing mass that depends on the source's label noise moves off the gold answer: default 0.02, SOC triage 0.1. Ordinal questions only smooth into neighbouring levels. CVSS severity uses a soft ordinal target from the raw base score, so a 6.9 puts real mass on "high" while a 5.0 sits fully in "medium". Multi-valued expert labels, such as a CVE whose techniques span several ATT&CK tactics, spread probability over every correct option. label is always the argmax; when several options are equally correct, it is the first of them, so score those questions with soft accuracy (the meta column lists every correct answer).

Leak-free splits.

• Sources with a native split keep it (CIRCL test, deepset/jackhhao test, typed-decisions test, NIST document-level splits).
• Every other source is split by a stable hash of its entity, not its row: a CVE id, a parent ATT&CK technique, a normalised log template or a normalised text.
• An entity that appears in several workflows is moved to its most held-out split in all of them, so no CVE description is ever in train for one question and test for another. 183 rows were moved by this rule.

Benchmark decontamination. Every CVE id and description in CTI-Bench's root-cause (CTI-RCM) and severity (CTI-VSP) tasks is removed from all splits. 53 rows were removed. Rook can therefore be evaluated on CTI-Bench without contamination. CTI-Bench data is not included here.

Leakage guards inside the text.

• ATT&CK descriptions have their technique name and header removed.
• CVE text has CWE ids masked.
• KEV text has the ransomware field removed.
• NIST excerpts have topic names and publication numbers removed.
• Option order is shuffled per row for CWE and mitigation questions.

Laya compatibility. Every distinct question was rendered with Laya's own build_head, and no options collapsed under the 192- or 256-token head budget. A stratified sample was run through the exact preprocessing in Laya's fine-tuning notebook without a failure. Share of questions whose state is truncated:

Workflow	max_len 512	max_len 1024
alert_triage	35.3%	0.0%
attack_technique	3.3%	0.0%
cwe_root_cause	0.0%	0.0%
guidance_domain	31.7%	0.0%
kev_ransomware	0.7%	0.0%
mitigation_select	0.0%	0.0%
phishing_message	10.0%	0.0%
phishing_url	0.0%	0.0%
prompt_attack	2.3%	0.0%
security_incidents	0.0%	0.0%
vuln_exploitation	0.2%	0.0%
vuln_severity	4.0%	0.0%

Per-workflow notes

• alert_triage: 3154 events; malicious = analyst-confirmed only; <= 2 events per log template; template-level splits
• attack_technique: split by parent technique, so sub-techniques never straddle splits
• cwe_root_cause: 1 expert-assigned CWE + 4 distractors that are never another valid CWE for the CVE
• guidance_domain: one case per NIST excerpt; multi-domain control catalogs excluded; native document-level splits kept
• kev_ransomware: 361 known / 1365 unknown; the ransomware field is removed from the input
• mitigation_select: 1 MITRE mitigates relationship + 4 distractors that never mitigate that technique
• phishing_message: emails + SMS; balanced; de-duplicated on normalised text
• phishing_url: balanced; de-duplicated
• prompt_attack: both native test splits kept as test; validation carved from train
• security_incidents: copied unchanged; official train/test split preserved
• vuln_exploitation: 1206 CVEs carry an ATT&CK tactic question; native test split kept
• vuln_severity: train classes balanced ({2: 2500, 3: 2500, 1: 2500, 0: 2500}); validation/test keep the natural class mix

Sources and credits

Rook is a derivative compilation. All credit for the underlying data goes to these authors:

Dataset	License	Used for
nuhmanpk/threatground	Other: MITRE ATT&CK / CWE Terms of Use; CISA KEV & NVD public domain	ATT&CK technique → tactic and platform map, CWE root-cause and mitigation choices, KEV ransomware use
nuhmanpk/cybersecurity-controls-instructions	CC BY 4.0 (NIST source documents are US public domain)	NIST guidance excerpts → security-domain routing
CIRCL/vulnerability-scores	CC BY 4.0	vulnerability descriptions → CVSS severity (soft ordinal targets)
CIRCL/vulnerability-attack-techniques	CC BY 4.0	expert CVE → ATT&CK mappings (MITRE CTID) and CVSS vectors
witfoo/precinct6-cybersecurity	Apache 2.0	sanitised production SOC events → triage level
ealvaradob/phishing-dataset	Apache 2.0	emails, SMS and URLs → phishing detection
deepset/prompt-injections	Apache 2.0	prompt-injection detection
jackhhao/jailbreak-classification	Apache 2.0	jailbreak detection
LocalLLaMA/typed-decisions	Apache 2.0	security_incidents workflow copied unchanged; also the row format Rook follows
RISys-Lab/Benchmarks_CyberSec_CTI-Bench (CTI-Bench)	CC BY-SA 4.0	not redistributed: only used to remove its CVEs from Rook (decontamination)

Underlying public sources: MITRE ATT&CK® and CWE™ (© The MITRE Corporation, used under their Terms of Use; ATT&CK and CWE are trademarks of The MITRE Corporation), the CISA Known Exploited Vulnerabilities Catalog and NIST publications (US Government works). This product uses data from the NVD API but is not endorsed or certified by the NVD.

License. Each row keeps the license of its source (source_license column). Filter on it if you need a single-license subset. CC BY 4.0 and Apache 2.0 require attribution, which this card provides. Check the original terms before redistributing derivatives.

Limitations and responsible use

• Rule-derived SOC labels. alert_triage labels come from WitFoo's detection rules and incident correlation, not from independent ground truth. The "suspicious vs. malicious" boundary depends on context beyond a single event.
• Partial view of a CVE. CVSS severity and ATT&CK tactics are judged from a description alone. The full advisory can change the answer.
• Coverage. Phishing data is mostly English and dates from 2023–2024. Prompt-attack data reflects jailbreak styles up to its collection date.
• Defensive use only. The dataset contains attack prompts, phishing text and exploit descriptions, but no exploit code. Use it for detection, triage and education.
• Not a replacement for analysts. A Rook model is a fast first-pass triage aid. Gate actions on calibrated confidence and keep a human in the loop.

Citations

If you use this dataset, please cite the sources it builds on:

@misc{bonhomme2025vlai,
  title={VLAI: A RoBERTa-Based Model for Automated Vulnerability Severity Classification},
  author={C{\'e}dric Bonhomme and Alexandre Dulaunoy}, year={2025}, eprint={2507.03607}, archivePrefix={arXiv}}

@misc{circl2026cveattack,
  title={Mapping CVEs to MITRE ATT\&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion},
  author={{Computer Incident Response Center Luxembourg (CIRCL)}}, year={2026}, eprint={2607.25572}, archivePrefix={arXiv},
  url={https://huggingface.co/papers/2607.25572}}

@dataset{witfoo_precinct6_2025,
  title={WitFoo Precinct6 Cybersecurity Dataset: Labeled Provenance Graphs and Signal Logs from Production SOC Operations},
  author={WitFoo, Inc.}, year={2025}, url={https://huggingface.co/datasets/witfoo/precinct6-cybersecurity}}

@inproceedings{alam2024ctibench,
  title={{CTIB}ench: A Benchmark for Evaluating {LLM}s in Cyber Threat Intelligence},
  author={Md Tanvirul Alam and Dipkamal Bhusal and Le Nguyen and Nidhi Rastogi},
  booktitle={NeurIPS 2024 Datasets and Benchmarks Track}, year={2024}, url={https://openreview.net/forum?id=iJAOpsXo2I}}

Author

Built by Nuhman PK: 🤗 Hugging Face · GitHub · LinkedIn · X