Meddies/meddies-pii

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Mar 4, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Meddies/meddies-pii is a causal language model developed by Meddies for multilingual PII extraction, fine-tuned from LiquidAI/LFM2-350M. It specializes in extracting 7 normalized PII entity families across 17 languages, returning structured JSON from clinical and administrative text. This model is designed for privacy and healthcare AI teams needing stable extraction behavior across varying languages and document formats, with a small footprint suitable for consumer GPUs and browser deployment.

Loading preview...

Meddies PII: Multilingual PII Extraction Model

Meddies PII is a causal language model specifically designed for structured PII (Personally Identifiable Information) extraction from multilingual clinical and administrative text. Developed by Meddies, this model addresses the critical need for robust PII handling in healthcare AI, aiming to provide stable extraction behavior regardless of language, document format, or runtime environment.

Key Capabilities

  • Multilingual Extraction: Supports PII extraction across 17 different languages.
  • Structured Output: Extracts 7 normalized PII entity families (e.g., address, human_name, email_address, phone_number, date, id_number, company_name) into deterministic JSON format.
  • Efficient Deployment: Features a small footprint, making it suitable for deployment on consumer GPUs and even in browser environments.
  • Foundation Model: Built upon the LiquidAI/LFM2-350M foundation model and fine-tuned with a specialized dataset.
  • Performance: Achieves an Entity F1 score of 0.8380 on its held-out test set, with strong performance on phone_number (0.9484 F1) and email_address (0.9252 F1).

Good Fits

  • Multilingual de-identification of clinical notes, discharge summaries, and administrative forms.
  • Browser or edge experiments where larger extraction models are impractical.
  • Establishing evaluation baselines for structured extraction in multilingual healthcare contexts.

Limitations

It's important to note that Meddies PII is an extractor, not a redaction or anonymization tool. It does not guarantee compliance with regulations like GDPR or HIPAA without local validation. Known weak spots include company_name extraction (0.3277 F1) and a small percentage of hallucinated values (around 1.3%). Nested entities and medical measurements (like dosages or blood pressure) are intentionally out of scope.

Finetunes
3 models