Aria AI — Cybersecurity
Collection
Evidence-linked cybersecurity review demos. Analyst workstations, not automatic enforcement. • 27 items • Updated
Datasets:
n_test int64 | split string | recall_at_3 float64 | recall_at_5 float64 | keyword_recall_at_3 float64 | citation_accuracy float64 | claim_coverage float64 | invented_quote_count int64 | abstention dict | stale_current_preferred float64 | conflict_flag_rate float64 | by_language dict | limitations list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
41 | question-family hold-out; documents shared | 0.97561 | 1 | 0.97561 | 1 | 1 | 0 | {
"precision": 0.8571428571428571,
"recall": 1,
"tp": 6,
"fp": 1,
"fn": 0,
"tn": 34
} | 0.875 | 1 | {
"en": {
"n": 23,
"recall_at_5": 1
},
"fa": {
"n": 18,
"recall_at_5": 1
}
} | [
"Synthetic runbooks. Lab recall is not operational IR quality.",
"Extractive answers only. Qwen/Kimi generation is not loaded.",
"NIST CSF organizes sections; this is not a product certification.",
"LLM-as-judge is not used as the only metric."
] |
IncidentRAG synthetic runbooks
58 short incident-response runbooks and 100 reference questions for the IncidentRAG demo. Seed 4.
This is fixture data (level 1). It does not prove operational incident-response quality. Organization dataset, collection, and static card are public. Live Gradio is created by scripts/publish.py.
Files
- data/documents.jsonl
- data/questions.jsonl
- data/splits.json
- data/schema.json
- data/evaluation.json
- data/sample_preview.json
- data/protocol.md
Splits
Split is by question family, not random rows of the same case. All documents are visible to both splits.
- development: phishing, ransomware (non-stale), account takeover, cross-cutting, some unanswerable
- test: data leak, cloud exposure, stale ransomware versions, IAM ownership conflict, remaining unanswerable
Question kinds: answerable, unanswerable, conflict, stale.
License
CC-BY-4.0. Keep the synthetic-data label. Do not treat laboratory host and ticket ids as live infrastructure.
- Downloads last month
- 46