eugenehp/privacy-filter-rs

OpenAI Privacy Filter — Rust/Burn Weights

Safetensors weights for openai/privacy-filter, packaged for inference with privacy-filter-rs (pure-Rust, Burn ML framework).

Contents

File	Size	Description
model.safetensors	2.6 GB	Model weights (bfloat16)
config.json	3 KB	HuggingFace model configuration
tokenizer.json	27 MB	BPE tokenizer (o200k_base)
tokenizer_config.json	234 B	Tokenizer metadata
viterbi_calibration.json	372 B	Viterbi decoder operating points

Model Details

• Architecture: Bidirectional transformer encoder with Sparse MoE
• Parameters: 1.5B total, ~50M active per token (top-4 of 128 experts)
• Hidden size: 640, Layers: 8, Heads: 14 Q / 2 KV (GQA)
• Context: 128,000 tokens (YaRN RoPE, sliding window 257)
• Output: 33 BIOES token classes over 8 privacy categories
• Dtype: bfloat16 (converted to f32 at load time by the Rust runtime)

Privacy Categories

1. account_number
2. private_address
3. private_date
4. private_email
5. private_person
6. private_phone
7. private_url
8. secret

Usage with privacy-filter-rs

# Clone the Rust project
git clone https://github.com/eugenehp/privacy-filter-rs
cd privacy-filter-rs

# Download weights into ./data (this repo)
# git clone https://huggingface.co/eugenehp/privacy-filter-rs data

# Run inference
cargo run --release -- -m data "My name is Alice Smith"

use privacy_filter_rs::{PrivacyFilterInference, backend::{B, Device}};
use std::path::Path;

let device = <Device as Default>::default();
let engine = PrivacyFilterInference::<B>::load(Path::new("data"), device)?;

let spans = engine.predict("My name is Alice Smith")?;
for s in &spans {
    println!("{}: {} (score: {:.4})", s.entity_group, s.word, s.score);
}
// private_person:  Alice Smith (score: 1.0000)

License

Apache 2.0 — same as the upstream openai/privacy-filter model.