CTI-Rationale

CTI-Rationale links Cyber Threat Intelligence (CTI) text to MITRE ATT&CK techniques and records why each mapping was made. Most datasets keep only the final technique label. Here every evidence span is tied to a technique and to a short rationale that justifies it, together with the technical primitives behind the mapping, the close techniques that were ruled out, and the type of reasoning used.

A correct label is not the same as a justified one, and a plain label does not let you tell the difference. The dataset can be used to train models that produce ATT&CK mappings grounded in evidence, and to check whether a model's reasoning is supported by the text rather than only whether its final answer is right.

At a glance

Reports	100 public threat reports
Publishers	8
ATT&CK version	Enterprise v18.1 (frozen)
Annotation unit	Paragraph (chunk)
Gold mappings	1,740 unique paragraph–technique pairs
Labeled paragraphs	1,394
Candidate negatives	3,100 unlabeled chunks
Identifiers covered	242 (102 parent techniques, 140 sub-techniques)
Independent annotator rows	A1: 1,393 · A2: 1,322
Annotation provenance	Human CTI analysts; no AI-generated rationales

Which files to use

Use the v2 files. They are the release analyzed in the accompanying paper and are the default configs.

File	Rows	Contents
cti_rationale_gold_v2.jsonl	1,740	Adjudicated gold mappings — start here
cti_rationale_all_v2.jsonl	4,455	A1 (1,393) + A2 (1,322) + gold (1,740) rows, for agreement studies
cti_rationale_chunks_v2.jsonl	4,494	Every curated chunk, with a labeled flag
report_split.csv	100	Frozen report-level split: 70 train / 15 validation / 15 test
schema_gold_v2.json	—	JSON Schema for a gold row
CHANGELOG_v2.json	—	Machine-readable record of the v1 → v2 repair

The original v1 files remain in the repository for provenance and are exposed as the *_v1 configs. They should not be used for new work — see the correction note below.

from datasets import load_dataset

gold = load_dataset("EbruResul/CTI-Rationale", "gold")            # 1,740 corrected mappings
allrows = load_dataset("EbruResul/CTI-Rationale", "all_annotators")
chunks = load_dataset("EbruResul/CTI-Rationale", "chunks")

v1 → v2 correction

The v1 gold file contained 1,741 rows but only 1,740 unique (report_id, chunk_id, technique_id) keys: two rows for unc4393-goes-gently-into-silentnight / P024 / T1080 described the same mapping.

v2 makes three repairs, all recorded in CHANGELOG_v2.json:

1. Deduplication. The two T1080 rows are merged deterministically — highest-confidence categorical values, the shortest evidence span, and the union of unique rationale clauses, primitives, and ruled-out alternatives.
2. confidence is an integer (was a quoted string).
3. primitives and ruled_out are arrays (were semicolon-delimited strings).

No annotation content was discarded.

Schema

Field	Type	Meaning
report_id, chunk_id	string	Paragraph identity; the pair is unique per chunk
source_url, vendor	string	Provenance of the source report
chunk_text	string	The paragraph the mapping was made from
technique_id, technique_name	string	ATT&CK v18.1 target and human-readable name
sub_technique_name	string	Sub-technique name where applicable
tactic	string	ATT&CK tactic
rationale	string	Free text. Human-authored justification
evidence_span	string	Shortest sufficient phrase copied from the paragraph
evidence_type	enum	explicit or implicit
confidence	integer 1–5	Adjudicated confidence
reasoning_type	enum	TECHNICAL_PRIMITIVE, BEHAVIORAL, LEXICAL, CONTEXTUAL
primitives	array[string]	Concrete mechanisms, commands, paths, artifacts
ruled_out	array[string]	Plausible ATT&CK alternatives rejected during annotation
annotator	string	gold, a1, or a2 (in the all-annotators config)

The rationale is free text; the six fields around it are the structured support. Please describe it that way rather than as a "structured rationale".

Important: the unlabeled chunks are candidate negatives

The 3,100 chunks without a gold mapping have not been reviewed and confirmed to contain no ATT&CK behavior. Absence from the mapping table is not evidence of absence.

Do not use them to estimate specificity, false-positive rate, or full-corpus precision. A blinded 512-item review protocol for the held-out test reports exists in the paper's supplementary package, but the human judgments are not yet complete.

How to evaluate

• Split by report, not by paragraph. Use report_split.csv. Paragraphs from one incident report share vocabulary and malware names; a paragraph-level split leaks.
• Report technique accuracy and grounding separately. In the accompanying experiments, better label F1 did not produce better overlap with human evidence spans.
• Expect a long tail. 58 identifiers occur once; the 40 most frequent account for 58.5% of mappings. Report macro scores alongside micro.
• Freeze the ontology. Labels are ATT&CK Enterprise v18.1. If you migrate to a later version, publish the identifier mapping and keep the original v18.1 fields.

Licensing

The authors' annotations — evidence spans, technique mappings, rationales, primitives, ruled-out alternatives, reasoning types, and confidence scores — are released under CC BY 4.0.

Source-report text remains the property of the respective publishers. This dataset records paragraph excerpts and source URLs for research use and does not redistribute complete reports.

Citation

@article{resul2026ctirationale,
  title   = {CTI-Rationale: An Expert-Annotated Dataset of Evidence-Grounded Rationales
             for MITRE ATT\&CK Technique Mapping},
  author  = {Resul, Ebru and Bouruc, Petru-Liviu and Rughini{\c{s}}, R{\u{a}}zvan-Victor
             and V{\u{a}}duva, Jan-Alexandru and {\c{T}}urcanu, Dinu},
  journal = {Journal of Cybersecurity and Privacy},
  year    = {2026}
}