privacy-filter-mlx (int4)

MLX-converted, int4-quantized weights of openai/privacy-filter, packaged for use with PrivacyFilterKit — a Swift package that runs on-device PII detection on Apple platforms via MLX-Swift.

Bundle contents

File	Purpose
weights.safetensors	int4 affine-quantized weights (group_size=64). Embedding + classifier head kept full-precision.
tokenizer.json	Hugging Face tokenizer (copied verbatim from upstream).
tokenizer_config.json	Tokenizer config.
id2label.json	33-label BIOES table (8 entity types: account_number, private_address, private_date, private_email, private_person, private_phone, private_url, secret).
model_config.json	Architecture parameters consumed by the Swift runtime.
MANIFEST.json	SHA-256 hashes of every file in the bundle.

Architecture

• 8 transformer layers, hidden size 640, 14 attention heads (2 KV heads, GQA)
• 128 local experts, top-4 MoE routing
• 200 064 vocab, 131 072 max position embeddings, sliding-window attention (128)
• 33-label BIOES head; the Swift decoder derives a BIOES validity mask at runtime (no learned CRF transition matrix in the upstream checkpoint)

Usage (Swift)

import PrivacyFilterKit

let bundle = URL(fileURLWithPath: "/path/to/privacy-filter-int4-v0.1.0")
let filter = try await PrivacyFilter(source: .directory(bundle))
let entities = try await filter.detect(in: "Email me at jane@example.com")

See the PrivacyFilterKit README for the full API.

Conversion pipeline

The conversion was produced by the scripts in privacy-filter-swift/scripts/:

1. 01_download_hf.py — download the upstream checkpoint
2. 02_export_config.py — extract label table, tokenizer, normalized model config
3. 03_convert_mlx.py — rename keys, downcast to bf16, write MLX-friendly safetensors
4. 04_quantize_mlx.py — int4 affine quantization (embedding + classifier head full-precision)
5. 06_export_bundle.py — assemble bundle + MANIFEST + tar.gz archive

License

Apache 2.0, inherited from the upstream model. See LICENSE.

Modifications from upstream

This bundle is a derivative of openai/privacy-filter. Significant changes:

• Weights converted from PyTorch safetensors to MLX-format safetensors (key rename + bf16 cast).
• int4 affine-quantized (group_size=64). Embedding, classifier head, and any transition matrix are kept full-precision.
• Bundle adds model_config.json, id2label.json, and MANIFEST.json for the Swift runtime; no model logic is changed.

Credits

• Upstream model: openai/privacy-filter
• Swift runtime: PrivacyFilterKit
• Conversion runtime: MLX / MLX-Swift