Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains ANONYMIZED threat-actor messages collected for cyber-threat-intelligence research. Opsec and PII identifiers have been replaced with type-preserving fakes or entity tags, but the content concerns real criminal activity. Access is reviewed manually. By requesting access you agree to use the data for defensive security research only and to honor the CC-BY-SA-4.0 share-alike terms.
Log in or Sign Up to review the conditions and access this dataset content.
EYENET Incident Gold (anonymized)
Human-reviewed, multi-label gold set of threat-actor messages (forum posts,
Telegram/Matrix chatter) tagged with the kind of cyber incident or criminal
activity they describe. This is the training/eval corpus behind the companion
model 4nt11/eyenet-incident.
Gated, manual approval. The data is anonymized but concerns real criminal activity; access requests are reviewed by hand.
Anonymization
Every record is passed through EYENET's export scrubber before it leaves the operator:
- opsec/domain identifiers (@handles, domains, URLs, crypto) -> type-preserving fakes;
- remaining PII (emails, phones, locations, names) -> Presidio entity tags;
- credentials and combo/stealer-log lines -> killed;
- attachment filenames -> body-scrubbed, with content hashes, platform file-ids and long digit runs stripped.
Nothing is deleted; values are replaced, so register and structure (the label signal) survive. Crew names, target countries and "leaked by" tags are kept on purpose - they are the signal, not victim PII.
Schema
One JSON object per line (train.jsonl):
| field | type | notes |
|---|---|---|
text |
string | anonymized message body |
labels |
list[float] (16) | multi-label vector, order = label list below |
tier |
string | reviewed (human) / ai / rescued - filter to reviewed for the human-only slice |
has_attachment |
bool | whether the source message carried an attachment |
att_kinds |
list[string] | attachment kind tags |
att_files |
list[string] | scrubbed attachment filenames (extension kept) |
Label order (16)
defacement ddos_attack intrusion breach_dump
credentials stealer_logs iab_corporate crimeware_tooling
crime_aas telecom_abuse phishing_delivery fraud_ops
infra_resale recruiting alliance crew_ops
Intended use and limitations
- Intended: training/evaluating CTI triage classifiers; reproducing the companion model. A defensive research resource.
- Not intended: deanonymization attempts, targeting individuals, or any operational action against the subjects of the data.
- Limitations: weighted toward the languages and boards EYENET collects
most; rare labels (
alliance,crew_ops) have limited support.
- Downloads last month
- 36