gliner2-PII-basque-v2

Version 2 of tknika/gliner2-PII-basque, a Basque-adapted fine-tune of fastino/gliner2-privacy-filter-PII-multi (205M-parameter multilingual PII detection model, GLiNER2 architecture, mDeBERTa-v3-base encoder).

Compared to v1, this version is trained with a v2 synthetic replay dataset with realistic value distributions (name frequencies from public statistics, coherent street/town/postcode triples from OpenStreetMap) and three entity types specific to educational contexts: user name (LMS/forum handles, including @-mentions), personal url (personal blogs and profiles) and student id.

Training

Two data sources were combined (experience replay, to avoid catastrophic forgetting of the base model's PII capabilities):

• Basque NER: the nerc_id split of orai-nlp/basqueGLUE (2,842 sentences), mapped to the base model's types: person name, location, organization, miscellaneous.
• Synthetic PII: the train split (5,000 sentences) of tknika/pii-synthetic-basque-v2 — Basque, Spanish and French sentences covering the whole Basque Country (Araba, Bizkaia, Gipuzkoa, Nafarroa, Iparralde), 17 PII types, checksum-valid values (mod-23 national IDs, mod-97 IBANs, Luhn-valid cards, real phone formats).

Training used LoRA (r=16, alpha=32) on the boundary/task heads, merged into the base model after training (~13 MB adapter). This repository contains the merged standalone model, the original adapter (lora-adapter/) and the full training recipe (errezeta/).

Results

PII detection

On the eval split of pii-synthetic-basque-v2 (1,000 sentences, 17 types, disjoint from training by construction):

Model	Precision	Recall	Micro F1
Base (zero-shot)	0.647	0.742	0.691
v1 (gliner2-PII-basque)	0.761	0.862	0.808
This model (v2)	0.949	0.919	0.934

Per-type F1 of this model (types sorted by frequency; base / v1 shown for the educational types):

Type	F1		Type	F1
person name	0.957		personal url	1.000 (base 0.175, v1 0.727)
email	1.000		organization	0.726
phone number	0.992		location	0.000 ⚠
user name	0.996 (base 0.864, v1 0.671)		student id	0.894 (v1 zero-shot 0.972)
date	1.000		bank account number	1.000
address	0.990		national id	0.771
city	0.723		age	1.000
date of birth	0.997		credit card number	1.000
zip code	0.997

Basque NER

BasqueGLUE validation sets (500 sentences per split), showing that the PII fine-tuning does not cause forgetting of the Basque NER capabilities:

Metric	Base (zero-shot)	This model
nerc_id/val person name F1	0.664	0.843
nerc_id/val micro F1	0.464	0.769
nerc_od/val (Wikipedia) person name F1	0.721	0.808
nerc_od/val micro F1	0.547	0.725

Usage

from gliner2 import AutoExtractor

model = AutoExtractor.from_pretrained("tknika/gliner2-PII-basque-v2")

text = ("Kaixo, @mikel_etxeberria naiz ikaslea, ikasle zk 123456B dut. "
        "Nire bloga https://mikeletxeberria.wordpress.com da eta "
        "helbidea Barakaldoko Nagusia kalea 42 da, 48901.")
result = model.extract_entities(
    text,
    ["person name", "email", "phone number", "national id",
     "bank account number", "credit card number", "date of birth",
     "date", "address", "city", "location", "zip code", "age",
     "organization", "user name", "personal url", "student id"],
)

Limitations

• The PII evaluation set is synthetic; real-world performance (messier text, ambiguous contexts) will be lower.
• The model systematically labels town names as city even where location would be expected (location F1 is 0 on the eval set).
• national id precision is 0.63 (over-prediction on other numeric identifiers).
• The synthetic replay data covers 17 of the 42 PII types of the base model; other types may degrade after fine-tuning.
• Research-quality evaluation; not production-tested.
• The license of the BasqueGLUE/EIEC training corpus should be verified before redistribution of derivatives beyond this model.