SatellaRedact-150M
lightweight PII detector for text
2 min readupdated
A compact model for detecting personally identifiable information (PII — names, addresses, phone numbers, card numbers, and IDs) in English text. Built by EphAsad on top of Answer.AI's ModernBERT-base encoder (149M parameters) and fine-tuned for NER (named entity recognition), it tags tokens using 39 BIO labels across 19 data categories. This lets you automatically redact sensitive data without spinning up heavy LLMs. Ideal for developers building local document processing pipelines and NLP privacy researchers.
Key features
- Detects 19 data types: first and last names, age, gender, street addresses, ZIP codes, phone numbers, emails, bank card numbers, passports, driver's licenses, and tax and insurance IDs.
- Runs on both GPUs and standard CPUs thanks to its compact footprint (149M parameters) and Safetensors format.
- Version v1.4 was fine-tuned on a mix of 41,000 examples, including synthetic hard-negatives, cutting false positives from 4.96% down to 0.47% compared to v1.3.
- Released under the permissive Apache 2.0 license, allowing use of the weights in your own projects with attribution.
What you can do
- Drop it into a local pre-processing pipeline to scrub sensitive data from documents before sending text to external cloud LLMs.
- Use it for semi-automated labeling and anonymization of English datasets with human-in-the-loop review.
- Adopt it as a lightweight baseline to benchmark compact token classifiers against large language models on PII detection tasks.
Example
from transformers import pipeline
detector = pipeline("token-classification", model="EphAsad/SatellaRedact-150M", aggregation_strategy="simple", device=-1)
text = "This fictional example refers to Alex Example in a demonstration record."
for item in detector(text):
print(item["entity_group"], item["start"], item["end"])
Read more
DetailsValidation metrics (2,500 examples): F1 — 95.39%, Precision — 94.83%, Recall — 95.96%.
On a full-text synthetic benchmark (670 documents, 570 entities) with custom post-processing, the model completely caught 97.37% of entities (555 out of 570), achieving a character-level Recall of 98.96%.
Training for v1.4 ran for 2 epochs on a single NVIDIA L4 (22 GiB VRAM) in 28 minutes and 45 seconds using OpenPII and synthetic data.
Limitations: English only; intended for research use, lacks GDPR certification, and requires human review in privacy-critical workflows.










Comments