# HolyCyber — Layer: `attack_knowledge`

A focused dataset for the `attack_knowledge` domain of the HolyCyber corpus, built on its own from that domain's authoritative sources only.

## Overview

- **Rows:** 2,925
- **Groups (leakage units):** 647
- **Entities in the domain:** 647
- **Splits:** test=280, train=2,367, val=278
- **Task types:** 10  |  **Subdomains:** 6
- **Row size (tiktoken cl100k, user+assistant):** min 2,060 / median 2,265 / mean 2,287 / p90 2,477 / max 2,648 tokens
- **Entity's own material:** median 6% of an answer's tokens; 2,734 rows are below 25% (the balance is the labelled reference appendix)
- **Total tokens:** 6,689,467  |  **Rows below the 2,000-token floor:** 0
- **Canary:** every row carries the decontamination canary

## Record format

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

```json
{"id": "6d408e38f3cb8dff", "row_key": "<domain>:<group_key>:<task_type>:<subdomain>:<part>",
 "domain": "attack_knowledge", "subdomain": "...",
 "task_type": "...", "difficulty": "intermediate", "group_key": "...",
 "source": "...", "license": "...", "split": "train", "canary": "...",
 "tokens": {"user": 180, "assistant": 2350, "total": 2530},
 "core_share": 0.41, "appendix": true,
 "references": [{"title": "...", "url": "..."}],
 "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` | this layer, and the slice of it the row covers |
| `task_type` | the question shape (detection, mitigation, mapping, ...) |
| `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

Topic sections used in this layer's answers (structural labels such as `Answer` or `Context` and the appendix are excluded):

| Section | Rows |
|---|---|
| Summary | 912 |
| Defensive use | 839 |
| Underlying weaknesses | 588 |
| Consequences | 584 |
| Classification | 541 |
| Definition | 539 |
| Options | 534 |
| Rationale | 534 |
| Reference | 534 |
| Chaining | 481 |
| Related attack patterns | 441 |
| The chain | 419 |

## How this layer is built

1. **Base cells** — the rows the source parsers emit for this domain (one per entity x task type x subdomain).
2. **Task shapes** — the same entity asked a different defensive question (detection, mitigation, framework mapping, response), built only from other sections of that entity's own record. No new facts are introduced.
3. **Reference pack** — a row still under the token floor is padded with compact descriptions of other entities *of this domain*, taken from this domain's own source text; the window rotates per row.

Every row therefore comes from this domain's sources only, clears the 2,000-token floor, and no two rows are the same.


## Task types

| Task type | Rows |
|---|---|
| relationship_reasoning | 953 |
| knowledge_explanation | 636 |
| identification | 534 |
| risk_assessment | 277 |
| chain_reasoning | 196 |
| definition | 161 |
| control_mapping | 72 |
| incident_response | 52 |
| mitigation_remediation | 41 |
| threat_attribution | 3 |

## Subdomains

| Subdomain | Rows |
|---|---|
| pattern_overview | 1,331 |
| pattern_chain | 558 |
| pattern_prerequisites | 421 |
| pattern_consequences | 329 |
| pattern_execution | 245 |
| pattern_mitigation | 41 |

## Sources

| Source | Rows |
|---|---|
| capec | 2,714 |
| mitre-attack | 211 |

## Quality controls

- **Exact duplicates removed:** 0
- **Near-duplicates removed:** 74 (MinHash-LSH, Jaccard >= 0.90, same task type only)
- **Quality gate:** 0 rejected ({})
- **Contaminated rows removed:** 0 ({})
- **Split integrity:** group-level by `group_key`; no group crosses a split

## Intended use

Supervised fine-tuning / evaluation of a defensive cybersecurity assistant specialised in this domain. All rows are defensive, detection or analysis oriented; none contains functional exploit code or weaponised prompts.

```python
from datasets import load_dataset
ds = load_dataset("json", data_files=dict(train="attack_knowledge/train.jsonl", validation="attack_knowledge/val.jsonl", test="attack_knowledge/test.jsonl"))
# rows whose answer is mostly the entity itself, not the appendix
core = ds["train"].filter(lambda r: r["core_share"] >= 0.5)
```

## Limitations

- 2,734 of 2,925 rows are more than 75% labelled reference appendix: this domain's own records are shorter than the 2,000-token floor, so the row is completed with other entities' records. Mask the appendix when fine-tuning if you only want the core answer.

- Paragraph-level reuse is measured in `research/AUDIT_REPORT.md`; rows are unique but their appendix material is shared.