Byrne-Anon

85M PII detector / anonymizer. BIOES tag per token, then redact the spans. SpikeWhale (Byrne) backbone plus a small token-classification head (hidden 640 → 33 BIOES labels).

Eight categories: private_person, private_email, private_phone, private_address, private_url, private_date, account_number, and secret (API keys, passwords, tokens, PINs).

Benchmark vs. openai/privacy-filter

I ran this against the 1.4B openai/privacy-filter token-classifier on 200 samples, per-token PII labels. ~1/16 the params, tagging tracks closely. Full 200-sample dump: eval_samples_200.jsonl.

Metric	Result
Per-token agreement (all tokens)	96.27%
Entity agreement (PII tokens only)	90.52%
Macro-F1 over PII classes	89.10

Per-category token-level scores:

Category	Precision	Recall	F1
private_email	94.6	99.6	97.1
private_url	95.0	99.3	97.1
account_number	86.7	94.0	90.2
private_address	85.3	95.7	90.2
private_date	94.3	84.7	89.2
private_person	90.9	78.5	84.2
secret	79.8	87.8	83.6
private_phone	98.2	69.1	81.2

Usage

from byrne_anon import ByrneAnon

anon = ByrneAnon(".")

anon.detect("Email me at sarah.j@example.com or call 415-555-0182.")
# [{'type': 'private_email', 'text': 'sarah.j@example.com', 'start': 11, 'end': 31},
#  {'type': 'private_phone', 'text': '415-555-0182',        'start': 39, 'end': 52}]

anon.redact("Ship to 1600 Pennsylvania Avenue NW, Washington DC 20500.")
# 'Ship to [PRIVATE_ADDRESS], [PRIVATE_ADDRESS].'

detect() returns entities with byte offsets. redact() replaces them with [TYPE] placeholders.

Files

File	Purpose
`model.safetensors`, `config.json`	SpikeWhale backbone weights + config
`pii_head.pt`	token-classification head (640 → 33 BIOES)
`pii_labels.json`	id → BIOES label map
`tokenizer.json`, `tokenizer_config.json`	byte-level SpikeTokenizer
`byrne_anon.py`	self-contained `detect()` / `redact()` API
`model_v2.py`, `config.py`, `spike_tokenizer.py`	SpikeWhale architecture + tokenizer

Limitations

English-centric evaluation.
Strongest on emails, URLs, accounts, addresses. Weaker: private_phone recall and secret precision. Unusual emails (+tags, very short, very long local parts) may be only partially tagged.
Span boundaries can include a leading space or clip by a token. Treat detections as approximate spans for redaction, not exact character forensics.
Custom architecture: load via the bundled byrne_anon.py (local modeling code, no remote code execution).

Citation

@misc{byrne2026byrneanon,
  title        = {Byrne-Anon: A Compact 85M PII Detector and Text Anonymizer},
  author       = {Byrne, Dean},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Quazim0t0/Byrne-Anon}},
}

License

Apache-2.0.