urchade/gliner_multi_pii-v1 continued-trained on PERSON detection for Indian names, using
228K synthetic name-in-context sentences.
Validated against 400 real, unseen Wikipedia biography sentences (not the synthetic templates
it trained on):
precision
recall
base model
63.6%
18.8%
this fine-tune
85.9%
99.3%
Other GLiNER entity types (e.g. DOB) are unaffected by this fine-tune — verified unchanged.
Part of hold-my-data, a fully local PII/secret
redaction tool. This model is loaded automatically by that project; it is not intended as a
general-purpose NER model outside that use case.
