CTI-Rationale

CTI-Rationale is an expert-annotated dataset that 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.

We built it because 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 to produce ATT&CK mappings with rationales that are tied to evidence, and to check whether a model's reasoning is grounded in the text and not only whether the final label is right.

At a glance

Reports	100 public threat reports
Vendors	8 security vendors
ATT&CK version	v18.1
Annotation unit	Paragraph (chunk)
Gold rows	1,741
Independent annotator rows	A1: 1,393 · A2: 1,322
Curated chunks (full corpus)	4,494 (1,394 labeled · 3,100 candidate negatives)
Unique techniques (gold)	242
Unique tactics (gold)	14
Language	English
Annotators	2 CTI-literate experts + reconciled gold

Files

CTI-Rationale/
├── cti_rationale_gold.jsonl     # main artifact: reconciled gold, 1,741 rows
├── cti_rationale_all.jsonl      # A1 + A2 + gold rows (annotator field), for IAA / per-annotator work
├── cti_rationale_chunks.jsonl   # full corpus: every curated chunk (4,494), labeled + 3,100 candidate negatives
├── <report>/                    # one folder per report (100 total)
│   ├── <report>_A1_final.csv    # annotator 1, independent
│   ├── <report>_A2_final.csv    # annotator 2, independent
│   ├── <report>_gold.csv        # reconciled gold for this report
│   ├── <report>_chunks.csv      # all curated chunks for this report (labeled + unlabeled candidate negatives)
│   └── <report>_url.txt         # source URL of the original report
└── README.md

Full corpus and candidate negatives

Besides the labeled mappings, we release every curated paragraph, whether or not it received a technique. cti_rationale_chunks.jsonl holds the whole corpus of 4,494 curated chunks, and each <report>_chunks.csv holds the same for one report. A chunk with labeled = false received no gold mapping. These 3,100 unlabeled chunks are the candidate negative pool: paragraphs that survived curation but were not mapped to a technique. They are candidate rather than verified negatives, so they can be used to build a verified negative set, but they should not be treated as guaranteed absences.

The gold is built by a reconciliation pass over the two independent annotations, A1 and A2. The annotators go through their disagreements together, using the official ATT&CK v18.1 documentation, and merge them into one agreed set. It is not a simple union or intersection. In the gold rationale field we keep both annotators' justifications for the same mapping, joined with ||, because two short justifications are more useful than one.

Schema

Each row is one (paragraph, ATT&CK technique) mapping.

Field	Description
report_id	Report slug (JSONL only)
annotator	a1, a2, or gold (JSONL only)
source_url	URL of the original report (JSONL only)
vendor	Publishing vendor domain (JSONL only)
chunk_id	Paragraph identifier within the report (e.g. P020)
chunk_text	The full paragraph text
tactic	ATT&CK tactic name
technique_id	ATT&CK technique or sub-technique ID (e.g. T1059.001)
technique_name	Technique name
sub_technique_name	Sub-technique name, if any
evidence_span	Minimal phrase in the paragraph that justifies the mapping
evidence_type	explicit (behavior stated directly) or implicit (inferred from context)
confidence	Annotator confidence, ordinal 1–5
reasoning_type	TECHNICAL_PRIMITIVE, BEHAVIORAL, LEXICAL, or CONTEXTUAL
primitives	;-separated technical primitives supporting the mapping (tools, commands, artifacts)
ruled_out	;-separated ATT&CK IDs of close alternatives that were rejected
rationale	Short natural-language justification (gold joins both annotators with ||)

Chunk files (cti_rationale_chunks.jsonl and <report>_chunks.csv)

The chunk files list every curated paragraph, not just the mapped ones.

Field	Description
report_id	Report slug (JSONL only; the folder name in the per-report CSV)
source_url	URL of the original report (JSONL only)
vendor	Publishing vendor domain (JSONL only)
chunk_id	Paragraph identifier within the report (e.g. P020)
chunk_text	The full paragraph text
labeled	true if the paragraph has at least one gold mapping, else false (candidate negative)
n_gold_mappings	Number of gold mappings for the paragraph (0 for candidate negatives)

Inter-annotator agreement

Agreement is computed on the matched pairs (mappings both annotators produced), using the metric appropriate to each field type.

Field	Metric	Score
technique_id	Cohen's κ	0.86
tactic	Cohen's κ	0.89
reasoning_type	Cohen's κ	0.76
evidence_type	Cohen's κ (raw agreement)	0.50 (0.98)
confidence	Krippendorff's α (interval)	0.78
evidence_span	Jaccard overlap	0.80
ruled_out	Jaccard overlap	0.69

evidence_type shows the kappa paradox: ~96% of evidence is explicit, so raw agreement is 0.98 but κ stays low on the imbalanced class. At paragraph level, the average overlap between the two annotators' technique sets is 0.42 — annotators do not always select the same number of techniques per paragraph, which is one reason a reconciled gold standard is built on top of the independent annotations.

Distributions (gold)

Tactics: Defense Evasion (395), Command and Control (242), Execution (183), Initial Access (149), Persistence (142), Discovery (133), Credential Access (132), Impact (100), Collection (97), Lateral Movement (66), Exfiltration (58), Privilege Escalation (28), Resource Development (14), Reconnaissance (2).

Reasoning types: TECHNICAL_PRIMITIVE (811), BEHAVIORAL (668), LEXICAL (227), CONTEXTUAL (35).

Evidence types: explicit (1,676), implicit (65).

Vendors: unit42.paloaltonetworks.com (15), cloud.google.com (15), blog.talosintelligence.com (15), kaspersky.com (12), cisa.gov (12), microsoft.com (11), welivesecurity.com (10), sophos.com (10).

Usage

from datasets import load_dataset

# reconciled gold (default)
ds = load_dataset("<your-namespace>/CTI-Rationale", split="train")

# all annotators (A1 + A2 + gold), e.g. for agreement studies
ds_all = load_dataset("<your-namespace>/CTI-Rationale", "all_annotators", split="train")

# full corpus incl. candidate negatives (labeled + unlabeled chunks)
ds_chunks = load_dataset("<your-namespace>/CTI-Rationale", "chunks", split="train")

Licensing and source text

The annotations in this dataset (evidence spans, technique mappings, rationales, primitives, ruled-out alternatives, reasoning types, confidence) are released under CC BY 4.0.

The source report text (chunk_text, evidence_span) is quoted from public threat reports that remain the property of their respective publishers (Palo Alto Unit 42, Google/Mandiant, Cisco Talos, Kaspersky, CISA, Microsoft, ESET WeLiveSecurity, Sophos). It is included for reproducibility and research use only; the source URL of every report is provided in the per-report _url.txt files and the source_url field. If you are a rights holder and want a report excluded, please open an issue.