SatellaRedact-150M v1.4

An English-language personal information detection model built on ModernBERT-base.

Author: EphAsad
Hugging Face: EphAsad
Architecture: ModernBERT-base (~149M parameters)
Task: Named Entity Recognition / PII detection
Labels: 39 BIO labels covering 19 entity types
Status: Experimental public research release

Overview

SatellaRedact-150M identifies potentially sensitive information in English text using a compact transformer encoder.

The model predicts token-level entity labels which can be converted into spans for downstream processing, including redaction, annotation, and document analysis.

It is designed as a lightweight alternative to larger language models for structured personal information detection.

Important: The model is not a certified anonymisation solution. It may miss sensitive entities or redact unrelated text. Human verification is necessary for privacy-critical applications.

Supported entity categories

AGE, BUILDINGNUM, CITY, CREDITCARDNUMBER, DATE, DRIVERLICENSENUM, EMAIL, GENDER, GIVENNAME, IDCARDNUM, PASSPORTNUM, SEX, SOCIALNUM, STREET, SURNAME, TAXNUM, TELEPHONENUM, TITLE, ZIPCODE

The model uses BIO tagging, with B- and I- labels for each category, plus the O label for non-entity tokens.

Architecture

SatellaRedact is based on ModernBERT-base, developed by Answer.AI.

ModernBERT is a bidirectional encoder architecture designed for efficient text representation and classification.

SatellaRedact adds a token-classification head trained to identify PII spans.

Property Value
Base model ModernBERT-base
Parameters Approximately 149M
Languages English
Model type Encoder / token classifier
Output BIO entity labels
Checkpoint format Safetensors
Inference CPU or GPU

Training methodology

SatellaRedact was developed through two main stages.

Stage 1 — v1.3

The initial model was fine-tuned from ModernBERT-base using English examples from the OpenPII dataset.

  • 40,000 training examples
  • 4,000 validation examples
  • 3 training epochs
  • English-language entity classification

The initial development also introduced corrections to span alignment and redaction coverage evaluation.

Stage 2 — v1.4

Version 1.4 continued fine-tuning from the saved v1.3 checkpoint, rather than starting again from ModernBERT-base.

The training mixture contained:

Source Examples
Original English OpenPII training data 26,000
Additional varied synthetic examples 12,000
Synthetic hard-negative examples 3,000
Total 41,000

Training configuration:

Parameter Value
Training windows 41,001
Validation windows 2,500
Epochs 2
Learning rate 1e-5
Training GPU NVIDIA L4
GPU memory Approximately 22 GiB available
Microbatch 4
Gradient accumulation 4
Effective training batch 16
Displayed training runtime 28 minutes 45 seconds

The synthetic augmentation targeted varied identifier formats, contextual personal information, and negative examples to improve generalisation and reduce unnecessary masking.

Evaluation

Token-level validation

Evaluated on 2,500 English validation examples.

Metric Score
Precision 94.83%
Recall 95.96%
F1 95.39%
Validation loss 0.02764

Selected category recalls:

Category Recall
Given name 89.08%
Surname 87.01%
Email 100.00%
Tax number 95.92%
ID card number 90.13%
Passport number 99.23%
Driver licence number 88.20%
Date 99.96%

These measurements use token/BIO labels. They should not be interpreted as guaranteed complete redaction of sensitive information.

Full-text synthetic challenge

A matched synthetic challenge containing 670 documents and 570 annotated entities was used to compare v1.3 against v1.4.

Metric v1.3 v1.4
Complete entity coverage 90.00% 97.37%
Sensitive character recall 95.65% 98.96%
Redaction character precision 60.90% 94.43%
Non-PII character false-positive rate 4.96% 0.47%

Version 1.4 completely covered 555 of 570 annotated entities.

On a separate 200-document OpenPII validation subset, the full-text evaluation reported 1,701 of 1,708 entities completely covered.

Evaluation clarification: Full-text measurements were produced using the project's custom inference and post-processing workflow, including pattern-based detection. The standard Transformers NER pipeline provided below does not include that post-processing, so its results may differ.

Evaluation limitations

The full-text challenge was synthetic, and the strict development release gate was not fully satisfied:

  • Complete entity coverage: 97.37%, below the 98% target.
  • Character recall: 98.96%, below the 99% target.
  • At least one category was below the 90% complete coverage target.

No independently established real-world privacy certification or external-domain performance claim is made.

Usage

Install the dependencies:

pip install -U transformers torch safetensors

Load the model using Hugging Face Transformers:

from transformers import pipeline

model_id = "EphAsad/SatellaRedact-150M"

detector = pipeline(
    task="token-classification",
    model=model_id,
    tokenizer=model_id,
    aggregation_strategy="simple",
    device=-1,
)

text = (
    "This fictional example refers to "
    "Alex Example in a demonstration record."
)

predictions = detector(text)

for item in predictions:
    print(
        item["entity_group"],
        item["start"],
        item["end"],
    )

The detector produces entity predictions. An application performing redaction should apply its own carefully evaluated span-merging, replacement, and review logic.

The standard pipeline does not reproduce the bespoke hybrid post-processing used for the full-text benchmarks.

Intended use

Potential applications include:

  • Experimental PII recognition
  • Research into privacy-aware NLP
  • Named entity extraction
  • Human-reviewed document annotation
  • Evaluation of lightweight PII detectors
  • Development of local text-processing pipelines

Limitations

  1. The model is primarily trained and evaluated for English.
  2. Many training and evaluation examples are synthetic.
  3. Unfamiliar identifier patterns can be missed.
  4. Predicted entity boundaries and labels may be incorrect.
  5. Some ordinary text may be classified as sensitive.
  6. High token-level F1 does not guarantee complete anonymisation.
  7. The model should not be used as the sole safeguard for confidential or regulated information.
  8. Privacy compliance, including GDPR or UK GDPR, cannot be inferred from benchmark results.
  9. Model weights are public, and memorisation of training content cannot be categorically excluded.

For sensitive applications, perform local inference when appropriate, minimise input data, validate outputs, and require human review.

Data sources and attribution

ModernBERT-base

Base model: answerdotai/ModernBERT-base.

ModernBERT-base is released under the Apache License 2.0.

OpenPII 1M

Dataset: ai4privacy/pii-masking-openpii-1m.

Attribution: Ai4Privacy / Ai Suisse SA. Copyright © 2026 Ai Suisse SA.

The dataset is licensed under Creative Commons Attribution 4.0 (CC BY 4.0). Its applicable attribution requirements remain in effect.

This model is an independent derivative development and is not endorsed by the dataset or base-model authors.

The model repository does not redistribute original training records.

License

SatellaRedact-150M model weights are released under the Apache License 2.0, with attribution to the upstream model and dataset as described above.

Citation

Please cite this model repository, the original ModernBERT model, and the OpenPII dataset when using SatellaRedact in research or derivative applications.

Model: https://huggingface.co/EphAsad/SatellaRedact-150M

Dataset: https://huggingface.co/datasets/ai4privacy/pii-masking-openpii-1m

Downloads last month
18
Safetensors
Model size
0.1B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for EphAsad/SatellaRedact-150M

Finetuned
(1,607)
this model

Dataset used to train EphAsad/SatellaRedact-150M