Benchmark Results (SPY)

Evaluated on the SPY benchmark (Savkin et al., 2025) with exact-match span-level metrics:

Key takeaways

• Highest F1 on both legal and medical domains.
• Best recall among GLiNER-based detectors (0.718 avg) — critical for redaction workflows where missed spans are data leaks.
• Consistent performance across domains (< 2-point F1 difference).

When to Use This Model

Redaction Example

def redact(text, labels, threshold=0.5):
    model = GLiNER2.from_pretrained("fastino/gliner2-pii-v1")
    result = model.extract_entities(
        text, labels, threshold=threshold,
        include_spans=True,
    )
    entities = result.get("entities", {})
    spans = []
    for label, values in entities.items():
        for value in values:
            start = text.find(value)
            if start != -1:
                spans.append((start, start + len(value), label))

    spans.sort(key=lambda s: s[0], reverse=True)
    redacted = text
    for start, end, label in spans:
        redacted = redacted[:start] + f"[{label.upper()}]" + redacted[end:]
    return redacted


text = "Please contact Maria Jensen at maria.jensen@example.dk or +45 20 12 34 56."
labels = ["person", "email", "phone_number"]
print(redact(text, labels))
# "Please contact [PERSON] at [EMAIL] or [PHONE_NUMBER]."

Training Details

Limitations

• Precision (0.35–0.37 on SPY) leaves room for improvement; the model tends to over-predict name entities, sometimes confusing common nouns, organisation names, and product names with personal names.
• Evaluated on a single benchmark (SPY) covering two domains. Broader multilingual and fine-grained evaluation is ongoing.
• Training data is fully synthetic and has not been validated by human annotators.
• Performance on non-European locales and scripts has not been measured.

Improving precision

For production use, consider:

• Per-label confidence thresholds (raise threshold for person / full_name)
• Dictionary-based filtering for common false positives
• Calibration on a small domain-specific development set

Citation

@misc{fastino2026gliner2pii,
  title   = {GLiNER2-PII: Multilingual PII Extraction via Synthetic Fine-Tuning},
  author  = {{Fastino AI Team}},
  year    = {2026},
  url     = {https://huggingface.co/fastino/gliner2-pii-v1}
}

Related work

@misc{zaratiana2026gliner2piimultilingualmodelpersonally,
      title={GLiNER2-PII: A Multilingual Model for Personally Identifiable Information Extraction}, 
      author={Urchade Zaratiana and Ash Lewis and George Hurn-Maloney},
      year={2026},
      eprint={2605.09973},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.09973}, 
}

@inproceedings{zaratiana-etal-2025-gliner2,
  title     = {GLiNER2: Schema-Driven Multi-Task Learning for Structured Information Extraction},
  author    = {Zaratiana, Urchade and Pasternak, Gil and Boyd, Oliver and Hurn-Maloney, George and Lewis, Ash},
  booktitle = {Proceedings of EMNLP 2025: System Demonstrations},
  year      = {2025}
}

@inproceedings{zaratiana-etal-2024-gliner,
  title     = {GLiNER: Generalist Model for Named Entity Recognition using Bidirectional Transformer},
  author    = {Zaratiana, Urchade and Tomeh, Nadi and Holat, Pierre and Charnois, Thierry},
  booktitle = {Proceedings of NAACL 2024},
  year      = {2024}
}

@misc{atreja2026pioneeragent,
  title  = {Pioneer Agent: Continual Improvement of Small Language Models in Production},
  author = {Atreja, Dhruv and White, Julia and Nayak, Nikhil and Zhang, Kelton and Princis, Henrijs and Hurn-Maloney, George and Lewis, Ash and Zaratiana, Urchade},
  year   = {2026},
  url    = {https://arxiv.org/abs/2604.09791}
}

License

Apache 2.0