QEV1.1-9B / DATASETS.md
querypieai's picture
Release self-contained QEV multitask model
0127113 verified
|
Raw History Blame Contribute Delete
1.21 kB
# Data and label policy
The training recipe combines the existing QEV general replay mix, official
multilingual DLP findings, and explicitly annotated Japanese vulnerability
examples. Every DLP record receives an `is_dlp` question plus type questions;
Japanese augmentation emphasizes full-/half-width forms, Unicode NFKC,
Japanese date-of-birth variants, driver's-license notation, phone/address
variants, passport IDs and table/line-wrap layouts.
The type map includes: person name, government ID, credit card, financial
account, password, medical record, email, phone, API key, auth token, device
ID, MAC address, birth date, IP address, driver's license, address, blood type,
passport, age, vehicle ID and IBAN. A record is `is_dlp=true` when source
findings or explicit vulnerability annotations are non-empty. Type labels come
only from those annotations; generated negative questions do not invent a
positive finding.
The released artifacts contain no raw records. The aggregate evaluation set is
6,281 rows (EN 1,065 / KO 1,064 / JA 4,152) and preserves no document text in
this package. Source URLs, licenses, revisions and processed hashes are listed
in `evaluation/EVALUATION_DATA_SOURCES.json`.