HolyCyber-Dataset

Layer	Rows	Groups	Train/Val/Test	Tasks	Min tok	Median tok	Core share	Token total
`threat_intelligence`	2,865	906	2,304/286/275	8	2,055	2,286	14%	6,584,918
`attack_knowledge`	2,925	647	2,367/278/280	10	2,060	2,265	6%	6,689,467
`vulnerability_management`	3,493	1,154	2,799/354/340	4	2,075	2,246	8%	7,875,961
`detection_engineering`	7,970	4,241	6,399/765/806	6	2,085	2,213	12%	17,964,031
`defense_engineering`	2,995	848	2,422/285/288	6	2,086	2,321	18%	6,962,624
`ai_ml_security`	4,602	1,716	3,687/479/436	7	2,051	2,888	100%	13,497,055
`secure_coding`	2,659	1,001	2,144/249/266	8	2,088	2,331	14%	6,272,640
`pentest`	2,846	785	2,302/285/259	8	2,050	2,316	18%	6,677,903
`incident_response`	2,993	625	2,407/296/290	7	2,088	2,379	12%	7,030,785
`network_security`	2,383	643	1,886/246/251	6	2,069	2,338	19%	5,713,268
`cloud_security`	2,254	590	1,804/232/218	7	2,076	2,313	18%	5,310,326
`malware_analysis`	2,105	1,073	1,684/223/198	4	2,081	2,348	16%	5,184,719
`identity_access`	2,310	623	1,870/234/206	6	2,080	2,330	19%	5,499,715
`cryptography`	2,101	579	1,688/228/185	6	2,064	2,873	100%	6,097,825
`governance_risk`	3,176	1,925	2,549/305/322	3	2,090	2,281	13%	7,737,824
`ml_security_methodology`	2,111	972	1,744/177/190	1	2,052	2,892	100%	6,311,722
`safety_boundary`	2,200	700	1,777/209/214	1	2,091	2,906	100%	6,653,709

Core share is the median percentage of an answer that is the entity's own material; the rest is a labelled reference appendix (## Reference: ...).

Each layer directory contains train.jsonl, val.jsonl, test.jsonl and a DATASET_CARD.md.

The global dataset (dataset/train.jsonl, val.jsonl, test.jsonl) is the concatenation of the layers with rows whose answer already occurs in another layer removed, so it holds a few fewer rows than the layer total and has no repeated answer text.

Every layer meets the 2,000-row target and the 2,000-token floor. Thin domains reach the row target by drawing on the wider public corpus (arXiv, IETF, OWASP, NIST); research/SOURCE_TABLE.md records what each layer rests on and research/AUDIT_REPORT.md reports the measured reuse of reference material.