Aria AI — Cybersecurity
Collection
Evidence-linked cybersecurity review demos. Analyst workstations, not automatic enforcement. • 27 items • Updated
Datasets:
split string | n int64 | entity_micro dict | entity_by_type dict | relation dict | citation_accuracy float64 | ungrounded_ioc_ratio float64 | fabricated_id_kept int64 | attack_mapped dict | baseline_regex dict | enhanced dict | brand_victim_rate float64 | limits string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
test | 12 | {
"precision": 1,
"recall": 1,
"f1": 1,
"tp": 32,
"pred": 32,
"gold": 32
} | {
"actor": {
"precision": 1,
"recall": 1,
"f1": 1,
"tp": 4,
"pred": 4,
"gold": 4
},
"domain": {
"precision": 1,
"recall": 1,
"f1": 1,
"tp": 8,
"pred": 8,
"gold": 8
},
"ipv4": {
"precision": 1,
"recall": 1,
"f1": 1,
"tp": 4,
"pred": 4,
"go... | {
"precision": 1,
"recall": 1,
"f1": 1,
"tp": 12,
"pred": 12,
"gold": 12
} | 1 | 0 | 0 | {
"precision": 1,
"recall": 1,
"f1": 1,
"tp": 8,
"pred": 8,
"gold": 8
} | {
"precision": 1,
"recall": 0.5,
"f1": 0.6667,
"tp": 16,
"pred": 16,
"gold": 32
} | {
"precision": 1,
"recall": 1,
"f1": 1,
"tp": 32,
"pred": 32,
"gold": 32
} | 0 | Synthetic fixtures. Lab precision is not operational CTI accuracy. |
ThreatIntel synthetic reports
32 synthetic English and Persian CTI notes for the ThreatIntel extraction demo. Seed 5.
Organization dataset and collection item are public. Live Gradio (AriaAICompany/threat-intel or alirezaaminzadeh/threat-intel) is created by scripts/publish.py after the daily Space-creation cap resets. This is fixture data (level 1). It does not prove operational extraction quality on real vendor reports. Reports are original laboratory text. They are not copies of copyrighted advisories.
Product: ThreatIntel · Collection: Aria AI — Cybersecurity · Site: aria-ai.ir
Files
- data/reports.jsonl
- data/reports/*.md
- data/splits.json
- data/schema.json
- data/protocol.md
- data/evaluation.json
- data/sample_preview.json
Splits
Split is by campaign, not random sentences of the same note.
- development: 5 campaigns × 4 variants = 20
- test: 3 campaigns × 4 variants = 12
- interactive: six
*-00fixtures used in the Space
Observed-intel, vendor-advisory hard negatives, brand-mention hard negatives, and insufficient fragments are all present.
Labels
claim_kind:observed,quote,author_recommendation,unknownvictim_status: brand mention defaults tonot_asserted- ATT&CK gold IDs are mapped only when a behavior phrase exists
- Rejected tokens such as
CVE-99-1must not be extracted
Held-out evaluation (12 test reports)
| Method | Precision | Recall | Notes |
|---|---|---|---|
| regex-ioc baseline | 1.00 | 0.50 | IOC shape only |
| gazetteer + relations | 1.00 | 1.00 | citation accuracy 1.00 |
| ATT&CK snapshot map | 1.00 | 1.00 | behavior phrase required |
Ungrounded IOC ratio 0. Fabricated IDs kept 0. Brand-as-victim rate 0. Lab scores are not operational CTI accuracy.
Recorded sample silverwave_loader_en-00: workflow needs_review (actor attribution stays uncertain), hash bef7faefd601d65d, 7 entities, 6 cited relations.
Persian
سیودو گزارش مصنوعی فارسی و انگلیسی برای استخراج شاخص و رابطه. تقسیم بر کمپین است. اعتبار شکلی بدخواهی نیست. نام برند قربانی نیست.
License
CC-BY-4.0 for the synthetic text. ATT&CK technique names in the lab snapshot are from MITRE ATT&CK (CC-BY). Keep the synthetic-data label. Do not treat documentation IPs or .example hosts as live infrastructure.
- Downloads last month
- 84