This is the final selected PIIBench model: a standard DeBERTa-v3-base token classifier trained directly on the prepared multi-source PII benchmark splits. It outperformed the source-conditioned hierarchical comparison model on the complete held-out experiment test split.

Paper

This model is released with the paper:

Fine-Tuning Over Architectural Complexity: Broad-Coverage PII Detection on PIIBench with DeBERTa
arXiv: https://arxiv.org/abs/2605.25816
Hugging Face Papers: https://huggingface.co/papers/2605.25816

This repository corresponds to the direct fine-tuned DeBERTa model reported as the final selected model in the paper.

Results

The reported evaluation uses the later prepared PIIBench experiment variant with 82 retained entity types and a held-out test split of 100,002 records. It is not the earlier 48-type Hub dataset release.

Held-Out Evaluation	Records	F1	Precision	Recall
Corrected heldout subset	5,000	0.6476	0.6300	0.6662
Complete experiment test split	100,002	0.6455	0.6277	0.6645

Full-test SHA-256: 65f8edc86399ba3f9e4ba44591d4583f9271f5d1df20e30a913305049559df77

Usage

This is a standard Transformers token-classification model:

python

from transformers import pipeline

pipe = pipeline(
    "token-classification",
    model="Pritesh-2711/piibench-deberta-base",
    aggregation_strategy="simple",
)
print(pipe("Contact me at jane@example.com."))

Related Resources

Dataset: Pritesh-2711/pii-bench
Source-conditioned hierarchical comparison model: Pritesh-2711/piibench-deberta-sch