deberta-v3-base_finetuned_bluegennx_run2.21_3e

Bluegennx: Sensitive Entity Detection for Indian Contexts

Developed by C4Scale, this model identifies personal, payment-related, and contextual entities in English-language text, with a focus on Indian names, addresses, and identifiers such as Aadhaar and PAN.

It is a fine-tuned version of Microsoft DeBERTa-v3-base for token classification, supporting 19 entity types. Applications can use its predictions to locate sensitive spans before applying their own masking, redaction, or replacement policies. The model differentiates in that it smartly recognizes entities in text. for instance. I am fifty five years old where fifty five is classified as AGE.

This repository contains an intermediate checkpoint from the Bluegennx.ai project, first published in June 2024.

Intended use

• Detect sensitive entities before text is sent to an LLM or another downstream service.
• Identify spans for masking or replacement in customer conversations, support tickets, and other business text.
• Support data preparation for model evaluation and fine-tuning, subject to validation of the resulting data.
• Evaluate Indian-context entity detection within an organization's own environment.

The model is a detector. Redaction, synthetic replacement, reversible mappings, access controls, and policy enforcement must be implemented by the surrounding application. Not every detected entity, such as a date or country, is sensitive in every context.

Quick start

Install the inference dependencies:

python -m pip install torch transformers sentencepiece

Run entity detection locally:

from transformers import AutoModelForTokenClassification, AutoTokenizer, pipeline

MODEL_ID = "C4Scale/deberta-v3-base_finetuned_bluegennx_run2.21_3e"

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_ID,
    use_fast=True,
    model_max_length=512,
)
model = AutoModelForTokenClassification.from_pretrained(MODEL_ID)

detector = pipeline(
    task="token-classification",
    model=model,
    tokenizer=tokenizer,
    aggregation_strategy="simple",
    device=-1,  # CPU; use device=0 for a compatible CUDA GPU.
)

# Fictional example; the email uses a reserved example domain.
text = "Please email Rajesh Kumar at [email protected] about the Chennai office."

for entity in detector(text, stride=64):
    start, end = entity["start"], entity["end"]
    print({
        "label": entity["entity_group"],
        "text": text[start:end],
        "start": start,
        "end": end,
        "score": float(entity["score"]),
    })

Results include the predicted entity type, score, and character offsets into the input text. The end offset is exclusive. Scores are model outputs, not calibrated guarantees that an entity is correct.

The example sets a 512-token input window and uses overlapping chunks for longer text. Chunking requires a fast tokenizer and an aggregation strategy. Validate entity boundaries and repeated mentions on your own inputs before using predictions for automatic redaction.

Supported entities and reported evaluation

The following results were reported by C4Scale for this checkpoint on an internal evaluation set. The evaluation data is not distributed, and these results have not been independently reproduced here. They do not establish performance on other datasets or deployment conditions.

Label spellings match the checkpoint, including AADHAR_CARD. Raw token predictions use BIO labels: B- marks the beginning of an entity, I- its continuation, and O text outside a labelled entity. The quick-start example groups token predictions into spans.

Data and release scope

C4Scale fine-tuned this checkpoint using an internally developed dataset focused on Indian entities and contexts. The training and evaluation datasets are proprietary and are not included in this release.

The public release provides model weights, tokenizer assets, model configuration, and this documentation. Internal data-generation, annotation, quality-filtering, and training pipelines are not included.

Feedback and collaboration

We welcome feedback on missed entities, false positives, and potential applications. Use the repository's Community tab and include the entity type, expected behaviour, and a fictional or appropriately sanitized example. Do not post actual personal information, payment details, or credentials.

For custom model evaluation, deployment, or API enquiries, contact C4Scale.

License

This public model checkpoint is released under the MIT license. The base model is Microsoft DeBERTa-v3-base.