Sinhala_codemix_pii_detection_model

Sinhala-English Codemix PII Detector

An XLM-RoBERTa token-classification model fine-tuned to detect personally identifiable information (PII) in informal Sinhala-English code-mixed text.

Model Details

Model Description

• Developed by: Ramosh99
• Funded by: University of Moratuwa, Department of Information Technology — undergraduate final-year research project (BSc IT, Level 4 Semester 7)
• Shared by: Ramosh99
• Model type: XLM-RoBERTa token classifier for BIO-style named-entity recognition
• Language(s) (NLP): Sinhala and English, including Sinhala-English code-mixed text
• License: MIT
• Finetuned from model: xlm-roberta-base

Model Sources

• Repository: Ramosh99/mixcode_pii_detector
• Paper: Undergraduate research thesis — PII Detection and Redaction in Low-Resource Code-Mixed Languages: A Sinhala-English Case Study, University of Moratuwa (2025)
• Demo: The accompanying Docker Space exposes a FastAPI inference service.

Uses

Direct Use

The model detects PII spans and assigns an entity type and confidence score. It is intended for Sinhala-English code-mixed conversational text and can also be tested on Sinhala-script input.

The research taxonomy contains these 22 entity types:

ACCOUNT_TYPE, ADDRESS, AGE, BANK_ACCOUNT, CARD_NUMBER, DATE, DESIGNATION, DISEASE, EMAIL, GENDER, LINK, LOCATION, NIC, ORG, PASSWORD, PERSON, PHONE, QUALIFICATION, RACE, RELIGION, TIME, USERNAME.

Downstream Use

Use the model as one component in a privacy pipeline for entity highlighting, masking, redaction, access-control review, or structured PII validation. Pair predictions with deterministic validators for formats such as phone numbers, NICs, email addresses, card numbers, and bank-account numbers.

Out-of-Scope Use

Do not use this model as the sole basis for legal, employment, financial, identity, or security decisions. It is not a complete de-identification guarantee, a calibrated risk classifier, or a replacement for human review. Do not assume that text without a prediction is safe to disclose.

Bias, Risks, and Limitations

The corpus is domain-specific and includes augmented examples. Real messages may contain spellings, transliterations, scripts, spacing, abbreviations, or PII formats that are not represented in training data. Entity frequencies are imbalanced, so rare sensitive classes may have lower recall than common classes. Sinhala morphology and script switching can also affect token boundaries and span extraction. Confidence values are model scores, not formally calibrated probabilities.

Recommendations

Evaluate the model on representative data from the intended deployment domain. Use confidence thresholds, format validation, audit logging, and human review for high-risk decisions. Minimize retention of submitted text and protect inference inputs because they may contain sensitive information.

How to Get Started with the Model

from transformers import pipeline

ner = pipeline(
    "ner",
    model="Ramosh99/mixcode_pii_detector_v3",
    aggregation_strategy="average",
)

text = "Mage name Kamal, man wada karanne MAS Holdings eke, phone 071-1234567"
for entity in ner(text):
    print(entity)

The accompanying Space also provides POST /predict and GET /health endpoints. The prediction request is:

{"text": "Mage name Kamal, man wada karanne MAS Holdings eke, phone 071-1234567"}

Training Details

Training Data

The model was fine-tuned on a custom annotated Sinhala-English code-mixed PII corpus. The data uses BIO tagging and includes augmented examples to improve coverage of low-frequency sensitive categories. A public dataset-card URL and full dataset release details were not specified in the available materials.

Training Procedure

The model was trained as a token-classification task using the Hugging Face Transformers ecosystem and XLM-RoBERTa's multilingual SentencePiece tokenizer. The project also developed character-to-subword alignment and span-recovery logic for code-mixed text, while the served API uses Transformers aggregation.

Preprocessing

Input text is tokenized with the XLM-RoBERTa tokenizer and aligned to BIO labels. Predictions are aggregated into entity spans, and the API removes B- and I- prefixes from returned entity_group values.

Training Hyperparameters

The available training materials record experiments using a learning rate of 3e-5, batch size 16, 5 epochs, weight decay 0.01, and warmup ratio 0.1. The exact configuration for the published checkpoint is not separately identified.

• Training regime: Mixed precision (fp16) on NVIDIA GPU with AdamW optimizer, linear warmup schedule, and early stopping on validation loss

Speeds, Sizes, Times

Training duration, throughput, checkpoint size, and total compute time were not recorded in the available model materials.

Evaluation

Testing Data, Factors & Metrics

Testing Data

One reported research benchmark used a template-aware split with 2,000 evaluation samples. A separate evaluation artifact covered 17 target entities; the available materials do not establish that both evaluations used identical splits or scopes.

Factors

The research analysis considers model architecture, entity type, entity frequency, and micro- versus macro-averaging. The attached confusion matrix shows a representative subset of BIO-label behavior rather than all 22 PII types.

Metrics

Precision measures the proportion of predicted entity spans that are correct. Recall measures the proportion of reference entity spans that are recovered. F1 is the harmonic mean of precision and recall. Micro-averaging weights entity instances by frequency, while macro-averaging gives each entity type equal weight and therefore exposes rare-label weaknesses.

Results

The associated XLM-RoBERTa comparison on the 2,000-sample benchmark reported:

A separate 17-entity evaluation reported:

Summary

The benchmark results show strong aggregate performance for the reported XLM-RoBERTa experiment. The lower macro recall in the separate 17-entity analysis indicates that aggregate scores can hide weaker performance on rare PII categories. These results are not a guarantee for every domain or entity format and should be reproduced on deployment-specific test data.

Evaluation Figures

<!--  -->

<!--  -->

<!--  -->

<!--  -->

<!--  -->

Model Examination

The evaluation artifacts examine entity-level precision, recall, F1 variation, confusion patterns, training and validation loss, and the relationship between entity frequency and F1 performance. No formal interpretability study was recorded.

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: NVIDIA GPU (T4, V100, A100), CPU, or Google TPU
• Hours used: Approximately 2–4 GPU hours per training run
• Cloud Provider: Google Colab / Kaggle (free-tier GPU)
• Compute Region: us-central1 (Google Colab default)
• Carbon Emitted: Estimated minimal; short training runs on shared GPU infrastructure

Technical Specifications

Model Architecture and Objective

The model is an XLM-RoBERTa token classifier fine-tuned with BIO-style labels for fine-grained PII named-entity recognition. XLM-RoBERTa provides multilingual representations and SentencePiece subword tokenization for mixed Sinhala and English input.

Compute Infrastructure

The exact compute infrastructure used for the published checkpoint was not recorded.

Hardware

Can be run in CPU/GPU/TPU

Software

The model uses the Hugging Face Transformers ecosystem. The deployed Space uses Python, FastAPI, Pydantic, and a Transformers NER pipeline.

Citation

If you use this model, please cite the project and record the exact model revision evaluated.

BibTeX:

@misc{ramosh99_mixcode_pii_2025,
  author       = {Ramosh99},
  title        = {Sinhala-English Codemix PII Detector: Fine-tuned XLM-RoBERTa for Code-Mixed PII Named-Entity Recognition},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Ramosh99/mixcode_pii_detector_v3}},
  note         = {Undergraduate FYP research, University of Moratuwa — BSc IT Level 4 Semester 7}
}

APA:

Ramosh99. (2025). Sinhala-English Codemix PII Detector: Fine-tuned XLM-RoBERTa for Code-Mixed PII Named-Entity Recognition [Model]. Hugging Face. https://huggingface.co/Ramosh99/mixcode_pii_detector_v3

Glossary

• PII: Personally identifiable information.
• NER: Named-entity recognition.
• BIO: Beginning, Inside, Outside tagging used to represent entity spans.
• Micro-F1: F1 calculated over all entity instances together.
• Macro-F1: The mean F1 across entity types.

More Information

The model is designed for Sinhala-English code-mixed PII detection rather than general-purpose NER. Predictions should be treated as evidence for a downstream privacy workflow, not as a definitive privacy classification.

Model Card Authors

The model card was prepared from the available Codemix PII detection project materials. Individual author names were not specified.

Model Card Contact

Model publisher: Ramosh99.