This repository provides a framework to automatically build, augment, and evaluate unstructured PII detection datasets using LLM-based template synthesis and augmentation pipelines. The goal of this project is to create a scalable and flexible dataset pipeline for Korean PII (Personally Identifiable Information) detection, leveraging LLMs (e.g., ChatGPT) to synthesize PII placeholders, augment data with semantic variations, and evaluate on transformer-based models.
- Generate natural language sentences with PII placeholders (e.g.,
{FULL_NAME},{EMAIL}). - Support diverse domains and contexts (finance, healthcare, government, etc.).
- Export generated templates in structured JSON format.
- Convert generated sentences into dataset structure with placeholders.
- LADAM-based augmentation: Layer-wise attention-driven augmentation.
- Faker-based injection: Replace placeholders with realistic synthetic PII values.
- BIO Tagging: Apply KLUE/Roberta tokenizer for token-level BIO labeling.
- Benchmark KLUE/Roberta series models on generated datasets.
- Provide precision/recall/F1 evaluation scripts.
- Enable direct comparison between augmentation strategies.
PII-bearing datasets are often scarce or privacy-restricted. This framework avoids handling real PII by generating synthetic datasets with placeholders and Faker-injected values.
- Generate unstructured sentences with placeholders using LLM prompts.
- Automatically construct dataset schema.
- Apply augmentation (LADAM, Faker, BIO tagging).
- Train and evaluate models (KLUE/Roberta).
- LLM-based Template Synthesis for diverse and realistic PII contexts.
- Augmentation Framework combining LADAM and Faker injection.
- Evaluation Suite for model comparison and reproducibility.