NERPA - Fine-Tuned GLiNER2 for PII Anonymisation

A fine-tuned GLiNER2 Large (340M params) model trained to detect Personally Identifiable Information (PII) in text. Built as a flexible, self-hosted replacement for AWS Comprehend at Overmind.

Fine-Tuning Details

• Base model: fastino/gliner2-large-v1 (DeBERTa v3 Large backbone, 340M params)
• Training data: 1,210 synthetic snippets generated with Gemini 3 Pro + Python Faker, each containing 2–4 PII entities
• Eval data: 300 held-out snippets (no template overlap with training)
• Strategy: Full weight fine-tuning with differential learning rates:
  • Encoder (DeBERTa v3): 1e-7
  • GLiNER-specific layers: 1e-6
• Batch size: 64
• Convergence: 175 steps

Why NERPA?

NERPA combines two technical advantages that commercial NER services like AWS Comprehend cannot offer:

1. Bi-Encoder Architecture for Zero-Shot Entity Detection

GLiNER2 is a bi-encoder that takes both text and entity label descriptions as input, rather than treating entity types as fixed output classes. This architectural difference means you can define arbitrary entity types at inference time without retraining:

# Standard PII entities
entities = detect_entities(model, text, entities={
    "PERSON_NAME": "Person name",
    "DATE_OF_BIRTH": "Date of birth",
    "EMAIL": "Email address",
})

# Add domain-specific entities on the fly
entities = detect_entities(model, text, entities={
    "PERSON_NAME": "Person name",
    "MEDICATION": "Drug or medication name",
    "DIAGNOSIS": "Medical condition or diagnosis",
    "LAB_VALUE": "Laboratory test result",
})

This isn't prompt engineering or few-shot learning. The model's bi-encoder architecture natively supports arbitrary entity schemas. Fine-tuning on PII improves precision on those specific types without degrading the zero-shot capability.

Example: Context-dependent entity distinction

text = """Last weekend, I visited Riverside Farm & Wildlife Park with my family. 
The kids were excited to see the tigers first—magnificent creatures pacing behind 
the reinforced glass. My daughter Sarah kept comparing them to our tabby cat at home, 
saying how similar their stripes looked, though obviously Mittens is much smaller and 
sleeps on our couch rather than prowling through artificial jungle habitats."""

entities = detect_entities(model, text, entities={
    "ZOO": "Animals in a zoo or wildlife park",
    "PET": "Pet animals owned by someone",
})

Output:

Last weekend, I visited Riverside Farm & Wildlife Park with my family. The kids were 
excited to see the [ZOO] first—magnificent creatures pacing behind the reinforced glass. 
My daughter Sarah kept comparing them to our [PET] at home, saying how similar their 
stripes looked, though obviously [PET] is much smaller and sleeps on our couch rather 
than prowling through artificial jungle habitats.

The model correctly distinguishes tigers (zoo animals) from the tabby cat and even the cat's name Mittens (pets) based purely on contextual cues. No retraining required.

2. Superior Performance on Standard PII

Fine-tuning GLiNER2 Large on 1,210 synthetic PII examples produced a model that outperforms AWS Comprehend on standard entity detection:

Model	Micro-Precision	Micro-Recall
AWS Comprehend	0.90	0.94
GLiNER2 Large (off-the-shelf)	0.84	0.89
NERPA (this model)	0.93	0.90

NERPA achieves 3% higher precision than AWS Comprehend while maintaining comparable recall. The fine-tuning also enables fine-grained date disambiguation (DATE_OF_BIRTH vs DATE_TIME), which AWS Comprehend cannot do without custom model training.

The Architecture Advantage

AWS Comprehend treats entity types as fixed classification targets. Adding a new entity type requires:

1. Annotating thousands of examples
2. Training a custom model
3. Paying for model hosting
4. Managing model versioning

NERPA's bi-encoder architecture makes entity types a runtime parameter. Adding new entities is a single line of code.

Pre-Optimised PII Entity Types

NERPA is fine-tuned on these entity types (but you can add more at inference time):

Entity	Description
PERSON_NAME	Person name
DATE_OF_BIRTH	Date of birth
DATE_TIME	Generic date and time
EMAIL	Email address
PHONE	Phone numbers
LOCATION	Address, city, country, postcode, street
AGE	Age of a person
BUSINESS_NAME	Business name
USERNAME	Username
URL	Any URL
BANK_ACCOUNT_DETAILS	IBAN, SWIFT, routing numbers, etc.
CARD_DETAILS	Card number, CVV, expiration
DIGITAL_KEYS	Passwords, PINs, API keys
PERSONAL_ID_NUMBERS	Passport, driving licence, tax IDs
TECHNICAL_ID_NUMBERS	IP/MAC addresses, serial numbers
VEHICLE_ID_NUMBERS	License plates, VINs

Quick Start

Install dependencies

pip install gliner2 torch

Anonymise text (CLI)

# Inline text
python anonymise.py "Dear John Smith, born 15/03/1990. Contact: john@acme.com"

# From file
python anonymise.py --file input.txt --output anonymised.txt

# Show detected entities
python anonymise.py --show-entities "Call me at 020-7946-0958, my IBAN is GB29NWBK60161331926819."

Use in Python

from anonymise import load_model, detect_entities, anonymise

model = load_model(".")  # path to this repo

text = (
    "Dear John Smith, your appointment is on 2025-03-15. "
    "Your date of birth (15/03/1990) has been verified. "
    "Please contact support at help@acme.com or call 020-7946-0958. "
)

entities = detect_entities(model, text)
print(anonymise(text, entities))

Output:

Dear [PERSON_NAME], your appointment is on [DATE_TIME].
Your date of birth ([DATE_OF_BIRTH]) has been verified.
Please contact support at [EMAIL] or call [PHONE].

Entity detection only

If you just need the raw entity offsets (e.g. for your own replacement logic):

entities = detect_entities(model, text)
for e in entities:
    print(f'{e["type"]:25s} [{e["start"]}:{e["end"]}] score={e["score"]:.2f}  "{text[e["start"]:e["end"]]}"')

PERSON_NAME               [5:15]  score=1.00  "John Smith"
DATE_TIME                 [40:50] score=1.00  "2025-03-15"
DATE_OF_BIRTH             [72:82] score=1.00  "15/03/1990"
EMAIL                     [129:142] score=1.00  "help@acme.com"
PHONE                     [151:164] score=1.00  "020-7946-0958"
BANK_ACCOUNT_DETAILS      [187:209] score=1.00  "GB29NWBK60161331926819"

Detect a subset of entities

entities = detect_entities(model, text, entities={
    "PERSON_NAME": "Person name",
    "EMAIL": "Email",
})

Custom entities

You can detect additional entity types beyond the built-in PII set. The model's zero-shot capability means any label + description pair will work — your custom entities are detected and anonymised alongside the fine-tuned ones.

CLI — use --extra-entities / -e:

python anonymise.py -e PRODUCT="Product name" -e SKILL="Professional skill" \
    "John Smith is a senior Python developer who bought a MacBook Pro."

Output:

[PERSON_NAME] is a senior [SKILL] developer who bought a [PRODUCT].

Python:

from anonymise import load_model, detect_entities, anonymise, PII_ENTITIES

model = load_model(".")

custom_entities = {
    **PII_ENTITIES,
    "PRODUCT": "Product name",
    "SKILL": "Professional skill",
}

text = "John Smith is a senior Python developer who bought a MacBook Pro."
entities = detect_entities(model, text, entities=custom_entities)
print(anonymise(text, entities))

How It Works

The inference pipeline in anonymise.py:

1. Chunking — Long texts are split into 3000-character chunks with 100-char overlap to stay within the model's context window. Specific chunk size can be varied since DeBERTa-v3 (underlying encoder) uses relative position encoding. We found that this size works as well as smaller ones.
2. Batch prediction — Chunks are fed through GLiNER2.batch_extract_entities() with include_spans=True to get character-level offsets.
3. Date disambiguation — Both DATE_TIME and DATE_OF_BIRTH are always detected together so the model can choose the best label per span.
4. De-duplication — Overlapping detections from chunk boundaries are merged, keeping the highest-confidence label for each position.
5. Replacement — Detected spans are replaced right-to-left with [ENTITY_TYPE] placeholders.

Notes

• Confidence threshold: Default is 0.25. The model sometimes tends to be conservative, so a lower threshold works well for high recall.
• GLiNER2 version: Requires gliner2>=1.2.4. Earlier versions had a bug where entity character offsets mapped to token positions instead of character positions; this is fixed in 1.2.4+.
• Device: Automatically uses CUDA > MPS > CPU.

Acknowledgements

This model is a fine-tuned version of GLiNER2 Large by Fastino AI. We thank the GLiNER2 authors for making their model and library openly available.

Citation

If you use NERPA, please cite both this model and the original GLiNER2 paper:

@misc{nerpa2025,
  title={NERPA: Fine-Tuned GLiNER2 for PII Anonymisation},
  author={Akhat Rakishev},
  year={2025},
  url={https://huggingface.co/OvermindLab/nerpa},
}

@misc{zaratiana2025gliner2efficientmultitaskinformation,
  title={GLiNER2: An Efficient Multi-Task Information Extraction System with Schema-Driven Interface},
  author={Urchade Zaratiana and Gil Pasternak and Oliver Boyd and George Hurn-Maloney and Ash Lewis},
  year={2025},
  eprint={2507.18546},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2507.18546},
}

Built by Akhat Rakishev at Overmind.

Overmind is infrastructure for end-to-end agent optimisation. Learn more at overmindlab.ai.