Meddies/meddies-pii

Meddies PII — Multilingual PII Extraction Model

A multilingual PII extractor for teams that need structured JSON from clinical and administrative text.

This is a research artifact for privacy and healthcare AI teams. It is not medical advice, not a redaction tool, and not a substitute for local validation before any clinical deployment, compliance workflow, or high-stakes privacy claim. If you want to use this model in commercial work, please contact us at contact@meddies-ai.com.

Why this model

PII handling is a load-bearing constraint in healthcare AI.

A model can sound clinically useful and still be unsafe if it leaks names, identifiers, phone numbers, email addresses, or addresses. Traditional NER pipelines also create friction: token alignment bugs, language-specific span normalization, and brittle post-processing when the document format shifts. Meddies PII is built for that problem. Give it raw multilingual text in chat format, and it returns normalized JSON keyed by the target entity families.

The goal is simple: keep extraction behavior stable when the language, document format, or runtime changes.

What this model does

Meddies PII is a causal language model used as a structured PII extractor.

Capabilities:

• multilingual extraction across 17 languages
• 7 normalized PII entity families
• deterministic JSON-friendly prompting
• a small enough footprint for consumer GPUs and browser deployment

Out of scope:

• automatic redaction or anonymization
• nested-entity reasoning
• adversarial hardening against evasive inputs

Quick start

Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Meddies/meddies-pii",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Meddies/meddies-pii")

messages = [
    {
        "role": "system",
        "content": "Extract <address>, <company_name>, <email_address>, <human_name>, <phone_number>, <id_number>, <date>",
    },
    {
        "role": "user",
        "content": "Patient John Smith, DOB 03/15/1985, was admitted to Mercy General Hospital. Contact: john.smith@email.com, (555) 123-4567. Address: 742 Evergreen Terrace, Springfield, IL 62704.",
    },
]

input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

output_ids = model.generate(
    input_ids,
    max_new_tokens=256,
    temperature=0.0,
    do_sample=False,
)

response = tokenizer.decode(
    output_ids[0][input_ids.shape[-1]:],
    skip_special_tokens=True,
)
print(response)

Expected output

{
  "human_name": ["John Smith"],
  "date": ["03/15/1985"],
  "company_name": ["Mercy General Hospital"],
  "email_address": ["john.smith@email.com"],
  "phone_number": ["(555) 123-4567"],
  "address": ["742 Evergreen Terrace, Springfield, IL 62704"]
}

The bundled chat_template.jinja now defaults to the full 7-label schema. Passing an explicit system prompt is still the safest way to keep extraction keys tight for your exact workflow.

vLLM

from vllm import LLM, SamplingParams

llm = LLM(model="Meddies/meddies-pii", dtype="bfloat16")
sampling = SamplingParams(temperature=0.0, max_tokens=512)

messages = [
    {
        "role": "system",
        "content": "Extract <address>, <company_name>, <email_address>, <human_name>, <phone_number>, <id_number>, <date>",
    },
    {
        "role": "user",
        "content": "Dr. Nguyen Van An, SĐT: 0912-345-678, email: an.nguyen@benhvien.vn",
    },
]

output = llm.chat(messages, sampling_params=sampling)
print(output[0].outputs[0].text)

Transformers.js (browser / Node.js)

import { pipeline } from "@huggingface/transformers";

const extractor = await pipeline("text-generation", "Meddies/meddies-pii-onnx", {
  dtype: "q4",
  device: "webgpu", // or "wasm" for broader compatibility
});

const messages = [
  {
    role: "system",
    content: "Extract <address>, <company_name>, <email_address>, <human_name>, <phone_number>, <id_number>, <date>",
  },
  {
    role: "user",
    content: "Patient John Smith, DOB 03/15/1985, contact: john.smith@email.com",
  },
];

const output = await extractor(messages, {
  max_new_tokens: 512,
  do_sample: false,
  temperature: 0.0,
});

console.log(output[0].generated_text.at(-1).content);

ONNX Runtime (Python)

from optimum.onnxruntime import ORTModelForCausalLM
from transformers import AutoTokenizer

model = ORTModelForCausalLM.from_pretrained("Meddies/meddies-pii-onnx")
tokenizer = AutoTokenizer.from_pretrained("Meddies/meddies-pii-onnx")

Evaluation details

Figure 1. Overall metrics first, then entity and language slices in the Meddies visual system.

Metric	Dataset / split	Result	Notes
Entity F1	Meddies/meddies-pii / eval	0.8110	Mixed-language validation slice
Precision	Meddies/meddies-pii / eval	0.8112	Exact-match entity scoring
Recall	Meddies/meddies-pii / eval	0.8109	Exact-match entity scoring
Entity F1	Meddies/meddies-pii / test	0.8380	Held-out test slice
Precision	Meddies/meddies-pii / test	0.8116	Exact-match entity scoring
Recall	Meddies/meddies-pii / test	0.8663	Highest overall headline metric
Value hallucination	eval / test	1.31% / 1.35%	Generated entity values not found in the input

Evaluation uses entity-level set-based exact match on (value, label) pairs. That is stricter than token overlap and closer to the extraction behavior a downstream system actually consumes.

Per-entity performance (eval)

Entity type	F1	Reading
phone_number	0.9484	Strongest class; formatting regularity helps
email_address	0.9252	Also strong due to rigid surface form
date	0.8607	Solid despite multilingual date variation
id_number	0.8132	Usable, but depends on locale formatting
address	0.7952	Harder because boundary detection is messy
human_name	0.7587	Sensitive to naming style and nested context
company_name	0.3277	Known weak spot from label-definition mismatch

Per-language performance (eval)

Full language table

Language	F1
Malay	0.8588
Korean	0.8539
Japanese	0.8497
Chinese	0.8461
Vietnamese	0.8251
Filipino	0.8126
Indonesian	0.8079
Burmese	0.7851
Portuguese	0.7802
Tamil	0.7740
Spanish	0.7772
French	0.7623
English	0.7528
German	0.7376
Thai	0.7303
Russian	0.7117
Lao	0.7077

The spread is usable but not flat. The model holds together across 17 languages, but the lower-resource slices still lag the stronger East Asian and Southeast Asian sources.

How it was built

• Foundation model: LiquidAI/LFM2-350M
• Training dataset: Meddies/meddies-pii
• Browser / ONNX variant: Meddies/meddies-pii-onnx

Figure 2. LFM2 foundation → full SFT on multilingual PII extraction → GRPO alignment with extraction-specific rewards → exact-match evaluation on eval and test → Hub and ONNX release.

Full reward design and GRPO configuration now live in TRAINING.md.

Good fits

Use this model when you care about extracted values more than token-level tagging internals.

Good fits include multilingual de-identification of clinical notes, discharge summaries, admin forms, and mixed healthcare documents; browser or edge experiments where larger extractors are too heavy; and evaluation baselines for structured extraction across multilingual healthcare text.

Limits

This is an extractor. Treat it that way.

• It does not redact or anonymize source text for you.
• Good benchmark numbers do not prove GDPR, HIPAA, or local-regulation compliance on your data.
• company_name is the weakest class in the current release.
• Around 1.3% of generated values are hallucinated rather than copied from the input.
• Nested entities are out of scope.
• If you omit the explicit system message, the bundled chat template defaults to five labels, not seven.
• Inputs designed to evade detection can still break this model.
• Medical measurements such as blood pressure, labs, dosages, and ages are intentionally excluded from the target label set.

Feedback

Send us the failures.

The useful reports are concrete: broken quick-start paths, false positives on measurements, misses on localized identifiers, hallucinated values, browser-runtime regressions, or language slices that collapse on your documents.

You can find Meddies on Hugging Face at huggingface.co/Meddies and on the web at meddies-ai.com.

Collaboration and sponsorship

Meddies is building verifiable clinical intelligence and the infrastructure around it.

We are a small team. Compute and review time are still tight.

If this work matters to you—sponsorship, collaboration, clinician review, or a larger conversation about the Meddies vision—email us at contact@meddies-ai.com.

Citation

@misc{meddies-pii-2026,
  title={Meddies PII: Multilingual PII Extraction with GRPO},
  author={Meddies Team},
  year={2026},
  url={https://huggingface.co/Meddies/meddies-pii}
}