opensec-triage 0.3.1

50,000 synthetic English security-alert examples for disposition classification and counterfactual evaluation.

Each example pairs a security observation with contextual evidence and an expected disposition. The task is to classify the supplied evidence.

The dataset contains 10,000 groups that share the same primary observation, asset and identity while varying the evidence. Each group contains 5 examples with distinct disposition labels.

Dataset splits

Split	Examples	Family profiles	Groups
train	40,000	80	8,000
validation	5,000	10	1,000
test	5,000	10	1,000

Family profiles, counterfactual groups, assets and identities are disjoint across splits.

Labels

Label	Examples	Meaning
malicious	7,143	Correlated evidence supports unauthorised activity by a compromised identity or process.
suspicious	7,143	An unexplained deviation warrants investigation without confirmed compromise.
benign	7,143	The activity matches an established ordinary business workflow.
expected_admin	7,143	The action matches a verified operational change and authorised operator.
authorised_testing	7,143	The action matches a verified security assessment and its scope.
misconfiguration	7,143	A reproduced configuration fault explains the event and its resolution.
unknown	7,142	Missing or conflicting evidence prevents a reliable disposition.

Loading

The default configuration contains text and target. The research configuration contains the same examples with additional metadata.

From the dataset directory:

from datasets import load_dataset

data = load_dataset('json', data_files={
    'train': 'data/train.jsonl',
    'validation': 'data/validation.jsonl',
    'test': 'data/test.jsonl',
})

Use the corresponding files under metadata/ to load the research records.

Data files

Path	Contents
data/*.jsonl	Text and target for each split.
metadata/*.jsonl	Aligned research records for each split.
metadata/behaviours.json	Observable definitions and investigation context.
metadata/families.json	Business workflows, operating contexts and ownership guidance.
metadata/variants.json	Evidence patterns and disposition policies.
metadata/record_schema.json	Research field types and required fields.
metadata/catalog_metadata.json	Dataset provenance and interpretation boundaries.
labels.json	Label definitions and evidence requirements.
manifest.json	SHA-256 file checksums.

Research records align row-for-row with the corresponding text-and-target files. They are an alternative representation of the same examples.

behaviour, family_id and variant_id refer to the matching catalog IDs. counterfactual_group_id identifies related examples. Use the version and configuration fingerprint when comparing record IDs across releases.

Composition and provenance

The catalogs contain 100 behavior definitions, 100 family profiles and 100 evidence variants.

Examples combine authored observations, business contexts and evidence patterns. Family profiles describe reusable business scenarios; evidence variants describe combinations of decision evidence. These categories do not identify malware families or independently observed incidents.

All records are synthetic. No production incident corpus was ingested. Asset and principal identifiers are synthetic placeholders. Investigation guidance and decision policies are authored context rather than collected telemetry.

Training and evaluation

Use text as the model input and target as the label. Research fields such as variant IDs, evidence patterns, signals, rationale, evidence status and decision policies encode label information and must not be additional model inputs.

Report macro-F1, per-class precision and recall, confusion matrices and consistency across related examples. Check tokenization so that evidence at the end of each text is retained.

Limitations

Evidence templates and constituent business contexts are shared across splits. Template memorisation can inflate test scores; unique text and held-out identities do not establish semantic independence.

Labels follow authored evidence rules rather than independent incident adjudication. The dataset does not measure generalisation to unseen evidence phrasing or real-world incidents. Some combinations of observations and workflows require domain review.

Disposition labels do not provide calibrated severity, confidence, likelihood or difficulty estimates. Business context and timing alone do not establish authorisation.

Assess operational suitability separately with analyst-reviewed cases and independently phrased evidence.

License

Licensed under CC BY 4.0.