README.md

---
license: apache-2.0
base_model: OpenMed/privacy-filter-multilingual-v2
datasets:
  - ai4privacy/pii-masking-200k
  - ai4privacy/pii-masking-400k
  - ai4privacy/open-pii-masking-500k-ai4privacy
  - ai4privacy/pii-masking-openpii-1m
  - nvidia/Nemotron-PII
  - gretelai/gretel-pii-masking-en-v1
  - piimb/privy
pipeline_tag: token-classification
library_name: openmed
tags:
  - openmed
  - mlx
  - apple-silicon
  - token-classification
  - pii
  - de-identification
  - medical
  - clinical
  - privacy-filter
  - multilingual
  - bf16
  - full-precision
language:
  - ar
  - bn
  - de
  - en
  - es
  - fr
  - hi
  - it
  - ja
  - ko
  - nl
  - pt
  - te
  - tr
  - vi
  - zh
---

# OpenMed Privacy Filter Multilingual v2 - MLX BF16

A native [MLX](https://github.com/ml-explore/mlx) port of [`OpenMed/privacy-filter-multilingual-v2`](https://huggingface.co/OpenMed/privacy-filter-multilingual-v2) for Apple Silicon PII detection and de-identification with OpenMed. This is the unquantized BF16 reference artifact. For the 8-bit sibling, see [`OpenMed/privacy-filter-multilingual-v2-mlx-8bit`](https://huggingface.co/OpenMed/privacy-filter-multilingual-v2-mlx-8bit).

> Family at a glance:
> - PyTorch source: [`OpenMed/privacy-filter-multilingual-v2`](https://huggingface.co/OpenMed/privacy-filter-multilingual-v2)
> - MLX BF16 (this repo): Apple Silicon, full precision, `2.6 GiB` weights
> - MLX 8-bit: [`OpenMed/privacy-filter-multilingual-v2-mlx-8bit`](https://huggingface.co/OpenMed/privacy-filter-multilingual-v2-mlx-8bit) - Apple Silicon, `1.4 GiB` weights

## At a glance

- Source checkpoint: [`OpenMed/privacy-filter-multilingual-v2`](https://huggingface.co/OpenMed/privacy-filter-multilingual-v2)
- OpenMed MLX repo: [`OpenMed/privacy-filter-multilingual-v2-mlx`](https://huggingface.co/OpenMed/privacy-filter-multilingual-v2-mlx)
- Label schema: 54 fine-grained multilingual PII categories
- Output space: 217 BIOES classes (O plus B/I/E/S for each category)
- Languages: 16 languages from the source card: ar, bn, de, en, es, fr, hi, it, ja, ko, nl, pt, te, tr, vi, zh
- Weight format: `safetensors`
- Quantization: none (BF16 reference)

## Q8 sibling validation

The 8-bit sibling was compared against this BF16 artifact on 10 golden PII samples. Decoded entity spans matched across all samples. Average Q8/BF16 argmax agreement was 100.00% with average logit MAE 0.1902; average local forward time was 15.1 ms for BF16 vs 8.4 ms for Q8.

## What it does

This model is an MLX packaging of [`OpenMed/privacy-filter-multilingual-v2`](https://huggingface.co/OpenMed/privacy-filter-multilingual-v2), the second-generation multilingual checkpoint for fine-grained PII extraction across 16 languages. It uses OpenAI's Privacy Filter architecture and predicts 217 BIOES classes (O plus B/I/E/S for each category). The OpenMed `PrivacyFilterMLXPipeline` runs BIOES-aware Viterbi decoding so callers receive grouped spans instead of raw token tags.

Label coverage highlights:

- Identity: FIRSTNAME, MIDDLENAME, LASTNAME, AGE, GENDER, USERNAME, OCCUPATION, ORGANIZATION
- Contact and address: EMAIL, PHONE, URL, STREET, BUILDINGNUMBER, CITY, COUNTY, STATE, ZIPCODE
- Financial and crypto: BANKACCOUNT, IBAN, BIC, CREDITCARD, CVV, PIN, BITCOINADDRESS, ETHEREUMADDRESS
- Vehicle, digital, and auth: VIN, VRM, IPADDRESS, MACADDRESS, IMEI, PASSWORD
- Date and amount labels such as DATE, DATEOFBIRTH, TIME, AMOUNT, CURRENCY, and CURRENCYCODE

The full label map is included in `id2label.json`.

## Architecture

| Field | Value |
| --- | --- |
| Source model type | `openai_privacy_filter` |
| Source architecture | `OpenAIPrivacyFilterForTokenClassification` |
| Hidden size | 640 |
| Transformer layers | 8 |
| Attention | Grouped-query attention (14 query heads / 2 KV heads, head_dim=64) with attention sinks |
| FFN | Sparse Mixture-of-Experts - 128 experts, top-4 routing, SwiGLU |
| Position encoding | YARN-scaled RoPE (`rope_theta=150000`, factor=32) |
| Context length | 131,072 tokens (initial 4,096) |
| Tokenizer | `o200k_base` / tiktoken-compatible tokenizer assets, vocab 200,064 |
| Output head | Linear(640 -> 217) with bias |

## File set

| File | Size | Purpose |
| --- | --- | --- |
| `weights.safetensors` | 2.6 GiB | MLX weights |
| `config.json` | 17.6 KiB | Model and OpenMed MLX runtime config |
| `id2label.json` | 4.8 KiB | Numeric ID to BIOES label mapping |
| `openmed-mlx.json` | 0.7 KiB | OpenMed MLX artifact manifest |
| `tokenizer.json` | 27 MiB | Tokenizer asset kept with the artifact |
| `tokenizer_config.json` | 0.2 KiB | Tokenizer metadata |

The MLX runtime uses the tiktoken-compatible `o200k_base` tokenizer path. `tokenizer.json` and `tokenizer_config.json` are bundled so consumers can inspect the tokenizer assets and keep the artifact self-contained.

## Quick start

### With OpenMed

```bash
pip install -U "openmed[mlx]"
```

```python
from openmed import extract_pii, deidentify
from openmed.core import OpenMedConfig

model_name = "OpenMed/privacy-filter-multilingual-v2-mlx"
text = (
    "Patient Sarah Johnson (DOB 03/15/1985), MRN 4872910, "
    "phone 415-555-0123, email sarah.johnson@example.com."
)

result = extract_pii(
    text,
    model_name=model_name,
    config=OpenMedConfig(backend="mlx"),
)
for ent in result.entities:
    print(ent.label, ent.text, round(ent.confidence, 4))

masked = deidentify(
    text,
    method="mask",
    model_name=model_name,
    config=OpenMedConfig(backend="mlx"),
)
print(masked.deidentified_text)
```

For non-MLX hosts, use the source PyTorch checkpoint [`OpenMed/privacy-filter-multilingual-v2`](https://huggingface.co/OpenMed/privacy-filter-multilingual-v2).

### Direct MLX usage

```python
from huggingface_hub import snapshot_download
from openmed.mlx.inference import PrivacyFilterMLXPipeline

model_path = snapshot_download("OpenMed/privacy-filter-multilingual-v2-mlx")
pipe = PrivacyFilterMLXPipeline(model_path)

print(pipe("Email me at alice.smith@example.com after 5pm."))
```

### Loading from a local snapshot

```python
from openmed.mlx.models import load_model
import mlx.core as mx

model = load_model("/path/to/privacy-filter-multilingual-v2-mlx")
ids = mx.array([[1, 100, 200, 300]], dtype=mx.int32)
mask = mx.ones((1, 4), dtype=mx.bool_)
logits = model(ids, attention_mask=mask)
print(logits.shape)
```

## Hardware notes

- Designed for Apple Silicon with MLX.
- CPU inference may work, but GPU-backed MLX on M-series Macs is the intended runtime.
- The Python package path is `pip install -U "openmed[mlx]"`.

## Credits

This artifact builds on:

- [`OpenMed/privacy-filter-multilingual-v2`](https://huggingface.co/OpenMed/privacy-filter-multilingual-v2) by OpenMed
- [`openai/privacy-filter`](https://huggingface.co/openai/privacy-filter) and OpenAI's `opf` training/evaluation tooling
- The datasets listed in the model-card metadata above
- Apple's [MLX](https://github.com/ml-explore/mlx) framework

## License

Apache 2.0, matching the source checkpoint metadata.