# HolyCyber — Dataset Card

A large, fully grounded, decontaminated instruction dataset spanning cybersecurity.

## Overview

- **Rows:** 51,805
- **Groups (leakage units):** 12,125
- **Splits:** test=5,702, train=40,998, val=5,105
- **Domains:** 17  |  **Task types:** 19
- **Avg assistant length:** 11,001 chars
- **Row size (tiktoken cl100k, user+assistant):** min 2,050 / median 2,324 / mean 2,464 / p90 2,926 / max 6,790 tokens
- **Total tokens:** 127,643,325  |  **Rows below the 2,000-token floor:** 0
- **Seed:** 1337  |  **Canary:** `BENCHMARK DATA SHOULD NEVER APPEAR AS TRAINING DATA: HOLYCYBER-DECONTAM-CANARY-9f2c7a14`
- **Entity's own material:** median 16% of an answer's tokens; 32,399 rows are below 25%. The balance is the labelled reference appendix (`## Reference: ...`), not extra facts about the row's entity.

## Record format

One JSON object per line; `messages` is a three-turn chat (system / user / assistant).

```json
{"id": "6d408e38f3cb8dff", "row_key": "detection_engineering:attack:T1568:detection_engineering:detection_strategy:1",
 "domain": "detection_engineering",
 "subdomain": "detection_strategy", "task_type": "detection_engineering",
 "difficulty": "advanced", "group_key": "attack:T1568", "source": "mitre-attack",
 "license": "...", "split": "train", "canary": "...",
 "tokens": {"user": 180, "assistant": 2350, "total": 2530},
 "core_share": 0.41, "appendix": true,
 "references": [{"title": "...", "url": "https://..."}],
 "messages": [{"role": "system", "content": "..."},
              {"role": "user", "content": "..."},
              {"role": "assistant", "content": "## ...\n..."}]}
```

| Field | Meaning |
|---|---|
| `id` | content hash of user+assistant, stable across rebuilds of the same row |
| `row_key` | stable identity across rebuilds (`domain:group:task:subdomain:part`) |
| `domain` / `subdomain` | the layer, and the slice of it the row covers |
| `task_type` | the question shape (detection, mitigation, mapping, response, ...) |
| `difficulty` | derived from the task shape, not guessed |
| `group_key` | the source entity; the unit of the train/val/test split |
| `source` / `license` | the authoritative source, and its licence |
| `tokens` | measured `cl100k_base` counts for user, assistant and total |
| `core_share` | share of the answer that is the entity's own material (1 - appendix) |
| `appendix` | whether the row carries a `## Reference:` appendix |
| `references` | resolvable links for the entities cited in the answer |
| `canary` | decontamination marker |

## What a row contains

Section headings used in the answers (appendix excluded):

| Section | Rows |
|---|---|
| Summary | 26,919 |
| Detection strategies and analytics | 7,440 |
| Mitigations | 6,854 |
| Tracked users (groups, malware, tools) | 6,836 |
| Classification | 5,088 |
| Introduction | 3,499 |
| Platforms | 3,123 |
| Background | 2,807 |
| Associated weaknesses | 2,379 |
| Scope note | 2,283 |
| What it is | 2,281 |
| Detection logic | 2,249 |
| Engineering note | 2,249 |
| Supports technique(s) | 2,232 |
| Telemetry | 2,210 |

## Row design

The corpus is built as **one focused dataset per cybersecurity domain** (`dataset/layers/<domain>/`), each from that domain's own authoritative sources, and this global file is their concatenation.

Each row is a **self-contained, long-form briefing on one entity** (an ATT&CK technique/group/software, a CWE weakness, a CAPEC pattern, a KEV vulnerability, a Sigma rule, an ATLAS technique, a D3FEND control, or a section of a cited arXiv paper). A row is built in three steps:

1. **Base cell** — the parser output for that entity and that (task type, subdomain).
2. **Task shapes** — the same entity asked a *different* defensive question (detection, mitigation, framework mapping, response, adversary usage), built only from other sections of that entity's own record, so no new fact is introduced.
3. **Reference pack** — if the row is still short, it is topped up with compact descriptions of other entities *of the same domain*, taken from that domain's own source text, with a rotating window so two rows do not carry identical padding.

Rows are packed to a hard floor of 2,000 tokens, so **no row is below the floor**. Context is allocated **disjointly** across an entity's cells, so a record is never repeated inside a single entity, and the split is cut on `group_key` so an entity never appears in two splits.

## Construction

Sources are parsed deterministically; no facts are invented by a model.

| Source | Rows | Licence |
|---|---|---|
| mitre-attack | 27,221 | MITRE ATT&CK — free for defensive use |
| arxiv-corpus | 11,332 | arXiv preprints; cited per row |
| cisa-kev | 5,843 | CISA KEV — U.S. Government work (public domain) |
| cwe | 2,747 | CWE (MITRE) — attribution required |
| capec | 2,714 | CAPEC (MITRE) — attribution required |
| ietf-rfc | 645 | see source |
| atlas | 608 | MITRE ATLAS — free for defensive use |
| nist-publication | 432 | see source |
| owasp-cheatsheet | 175 | see source |
| research-corpus | 88 | Curated; cites public arXiv preprints |

## Quality controls

- **Exact duplicates removed:** 0
- **Near-duplicates removed:** 952 (MinHash-LSH: 5-word shingles, 112 perms, Jaccard >= 0.9; same task type only)
- **Decontamination:** token-25 overlap vs 27 benchmark files; 2,064 benchmark CVE IDs excluded; canary GUID emitted per row
- **Contaminated rows removed:** 0 ({})
- **Quality gate:** 315 rejected ({})
- **Split integrity:** group-level by `group_key`; no group appears in more than one split

## Domain coverage

| Domain | Rows | Subdomains covered |
|---|---|---|
| detection_engineering | 7,970 | 6 |
| ai_ml_security | 4,602 | 10 |
| vulnerability_management | 3,493 | 4 |
| governance_risk | 3,176 | 5 |
| defense_engineering | 2,986 | 3 |
| incident_response | 2,943 | 4 |
| attack_knowledge | 2,887 | 6 |
| threat_intelligence | 2,865 | 6 |
| pentest | 2,788 | 7 |
| secure_coding | 2,659 | 6 |
| network_security | 2,373 | 5 |
| identity_access | 2,304 | 4 |
| cloud_security | 2,244 | 5 |
| safety_boundary | 2,200 | 5 |
| ml_security_methodology | 2,111 | 6 |
| malware_analysis | 2,104 | 4 |
| cryptography | 2,100 | 5 |

**Known coverage gaps (subdomains not yet populated):**

- threat_intelligence: malware, procedure
- vulnerability_management: prioritisation, weakness_consequences, weakness_detection, weakness_mitigation
- detection_engineering: false_positive_analysis
- defense_engineering: architecture_control, hardening
- incident_response: eradication, lessons_learned, recovery
- network_security: exfiltration
- cloud_security: storage_exposure
- malware_analysis: aliases, static_indicators
- identity_access: kerberos, mfa_bypass

## Intended use

Supervised fine-tuning / instruction tuning of a defensive cybersecurity assistant (HolyCyber persona). All rows are defensive or detection/analysis oriented.

## Out of scope

No functional exploits, malware, weaponised jailbreak prompts, or operational attack tooling. Pentest rows are framed as authorised engagements; offensive concepts are conceptual and paired with detection/remediation.

## Limitations

Rows are unique; some paragraphs are not. The framework domains complete short entity records with a labelled reference appendix, so a sizeable share of their answer tokens is reused material from the same domain. `research/AUDIT_REPORT.md` measures the paragraph reuse, including how much of it crosses the train/test boundary, and `research/EXTERNAL_REVIEW.md` records how this compares with a public peer dataset.

No row has been reviewed individually by a domain expert: grounding bounds invention, it does not by itself prove correctness.