Fastino releases open-source GLiNER2-PII model
Fastino has open-sourced GLiNER2-PII, a 300M-parameter PII detector it says beats OpenAI's Privacy Filter on the SPY benchmark and supports custom redaction schemas without any retraining.
Fastino announced GLiNER2-PII, a 300 million parameter multilingual encoder model that detects and redacts personally identifiable information in unstructured text. It is available on the Hugging Face Hub under what the company calls a permissive license for research and production use.
Out of the box, the model recognizes 42 entity types in seven categories: personal identity, contact and location, government and tax IDs, banking and payment, digital identity, secrets and credentials, and sensitive dates. It is label-conditioned, so the target schema is supplied as an input rather than fixed in the weights. According to Fastino, that lets one checkpoint serve different redaction policies, such as keeping a card expiry while masking the CVV. Because it is an encoder, all labels are scored in a single forward pass and the output is deterministic.
Fastino contrasts this with OpenAI's Privacy Filter, which it describes as a 1.5B-parameter decoder checkpoint repurposed as a token classifier and limited to eight fixed entity types.
On the SPY benchmark, which has 200 legal Q&A and medical transcript documents labeled with seven PII types, GLiNER2-PII posted an average F1 of 0.471. The four comparison models scored 0.391 (NVIDIA GLiNER PII), 0.384 (urchade), 0.373 (OpenAI Privacy Filter), and 0.368 (Knowledgator). Fastino says it applied the same label mapping to all five systems. These results come from the vendor and have not been independently verified.
The company credits the gains to post-training on 4,910 synthetic annotated examples in seven languages. Its Fine-Tuning Agent generated the data.