How to use from the
Use from the
Transformers library
# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("token-classification", model="Tracer-ML/PIIX-560m")
# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForTokenClassification

tokenizer = AutoTokenizer.from_pretrained("Tracer-ML/PIIX-560m")
model = AutoModelForTokenClassification.from_pretrained("Tracer-ML/PIIX-560m", device_map="auto")
Quick Links

PIIX-560m

PIIX-560m is a multilingual model for detecting and masking personally identifiable information (PII) and secrets in text. It combines a fine-tuned XLM-RoBERTa-large token classifier (560M parameters) with a rule layer for structured identifiers such as card numbers, IBANs and API keys.

  • Developed by: Tracer
  • Model type: token classification with rule-based post-processing
  • Languages: English, Russian, German, French, Spanish, Italian, Dutch, Portuguese, Polish, Ukrainian, Turkish, Kazakh, Belarusian
  • License: Apache 2.0
  • Base model: FacebookAI/xlm-roberta-large

Usage

The full pipeline lives in piix.py (tagger, rules, type filter, whole-word masking). Download piix.py, rules.py and type_filter.py from this repository.

pip install torch transformers safetensors
from piix import Redactor

redactor = Redactor()  # loads Tracer-ML/PIIX-560m on CUDA, Apple GPU or CPU

redactor.redact("Hi, this is Maria Gonzalez from Acme Corp. Reach me at maria.g@acme.io or +1 415 555 0134.")
# 'Hi, this is [REDACTED] from [REDACTED] Reach me at [REDACTED] or [REDACTED]'

result = redactor.find(
    ["Contact jdoe_91, tel 415-555-0134"],
    types=["user name", "phone number"],
    recall=4,
)[0]
result["findings"]  # typed spans with character offsets
result["mask"]      # merged spans to hide

Command line:

echo "some text" | python piix.py --recall 4

Parameters

Parameter Description
recall Sensitivity. 0 (default) favours precision, 4 is balanced, 8 maximises coverage.
types Optional list of entity type names to keep, in plain English. If a name is not recognised, all types are kept and the name is listed in redactor.unrecognised_types.
use_rules Enable the rule layer for structured identifiers. Default True.
widen Extend every finding to whole words. Default True.

Redactor(adapter=...) loads a fine-tuned adapter (see Fine-tuning).

Documents of any length are supported through overlapping 512-token windows.

The reported results use the full pipeline. Loading the checkpoint as a plain token-classification pipeline skips the rule layer and masking logic and gives lower coverage.

Fine-tuning

PIIX-560m can be adapted to a specific labelling policy with a LoRA adapter trained on a few hundred hand-labelled documents. See FINETUNING.md.

pip install peft
python piix_finetune.py fit labelled.jsonl --out my_adapter
redactor = Redactor(adapter="my_adapter")
Labelled documents Gretel PII masking ai4privacy 300k (English) ai4privacy 300k (de, es, fr, it, nl)
0 (base model) 89.2 85.8 83.5
50 93.3 91.0 89.9
200 94.1 92.1 90.8
1,000 95.6 93.7 92.1

F1 on held-out documents after fine-tuning on documents from the same dataset. The base model row is measured with no types list.

Entity types

Group Types
People given name, family name, person, username
Organizations organization
Contact email, phone number, postal address, location
Identifiers account number, employee ID, national ID, passport, tax number, other ID numbers
Technical URL, IP address, MAC address, hostname
Secrets API keys, tokens, passwords, private keys, credentials in connection strings
Dates dates
Sensitive attributes age, gender, marital status, nationality, belief, medical, financial, HR record

Evaluation

Character-level precision, recall and F1 are computed on masked characters, ignoring entity type. Not fully hidden is the share of gold spans with at least one visible character after masking (lower is better).

PII Arena

Tracer evaluated the released PIIX-560m on 40 of the 41 frozen PII Arena v1.0.3 datasets. Each setting covered 40,147 records and 226,720 annotated spans. Arena’s strict scorer reproduced all 80 per-dataset results. Rows marked * use Arena’s published counts for the same 40 datasets.

Model Not fully hidden ↓ Missed ↓ Precision Recall F1
PIIX-560m (recall=8) 3.46% 2.16% 54.26 97.45 69.71
PII-Tracer + GLiNER2 + BardsAI + mmBERT ensemble* 5.61% 1.44% 53.2 96.6 68.6
PIIX-560m (default) 9.36% 6.55% 69.26 93.04 79.41
Fastino GLiNER2* 20.17% 13.22% 75.1 77.7 76.4
Perplexity PII-Tracer* 25.14% 21.00% 64.9 79.6 71.5
OpenAI Privacy Filter* 40.43% 34.72%

The submission requests evaluation on the remaining dataset, secrets-rules. Tracer’s measurements, adapter, and reproduction instructions

PrivacyBench+

PrivacyBench+: enterprise chat messages and documents, 1,877 records.

Model Precision Recall F1
PIIX-560m 96.5 97.2 96.9
Fastino GLiNER2.5-base 89.3 83.4 86.2
Perplexity PII-Tracer 86.1 72.9 79.0
OpenAI Privacy Filter 87.6 64.1 74.1

Additional benchmarks

F1, with the share of spans not fully hidden in parentheses.

Model Gretel PII masking (test) ai4privacy 300k (English) ai4privacy 300k (de, es, fr, it, nl)
Perplexity PII-Tracer 95.4 (1.47%) 95.5 (0.39%) 95.0 (0.65%)
Nym base 93.9 (12.84%) 91.4 (13.52%) 91.2 (13.18%)
PIIX-560m 92.1 (1.19%) 85.7 (5.99%) 83.5 (6.53%)
OpenAI Privacy Filter 90.2 (16.16%) 91.1 (21.11%) 90.6 (21.00%)

Gretel: 5,000 test documents. ai4privacy: 3,000 validation rows per column. All models evaluated with the PII Arena runner and the same scorer.

Speed

Hardware Chat message Document page
NVIDIA A100 26.5 ms 61.9 ms
Apple M5 Max 32 ms 214 ms
NVIDIA T4 48 ms 300 ms

Limitations

  • At the default setting the model masks all dates and organization names it finds. Use types to restrict the output when a narrower definition of PII is needed.
  • Higher recall values increase coverage and also increase over-masking.
  • Indirect identifiers, such as a product name that reveals an employer, are only partially covered.
  • Results on domains far from the training data may be lower than those of large language models.
  • Outputs can differ slightly across hardware because of mixed-precision inference.
  • The model does not guarantee anonymization. Review its output before relying on it for compliance.

Training

Trained on 31M tokens. The training data is not released.

License

Apache License 2.0.

Citation

@misc{piix560m2026,
  title  = {PIIX-560m},
  author = {Tracer},
  year   = {2026},
  url    = {https://huggingface.co/Tracer-ML/PIIX-560m}
}
Downloads last month
34
Safetensors
Model size
0.6B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Tracer-ML/PIIX-560m

Finetuned
(1,108)
this model