pii-removal-edge-deployment

End-to-end workflow for generating synthetic PII-containing healthcare text, fine-tuning a small language model (SLM) with LoRA adapters for PII removal / redaction, and preparing an optimized GGUF artifact for edge / resource-constrained deployment.

Project Overview

This repository demonstrates a three-stage pipeline:

1. Synthetic Data Generation (notebooks/step_1_synthetic_data_generation.ipynb): Creates structured healthcare-style text with embedded PII (names, dates, MRNs, etc.).
2. Model Fine-Tuning (notebooks/step_2_slm_finetuning.ipynb): Applies LoRA to adapt a small base model for PII detection / transformation.
3. Edge Deployment (notebooks/step_3_slm_edge_deployment.ipynb): Exports / converts the model to GGUF and provides patterns for lightweight inference.

Repository Structure

data/
  synthetic_data/           # Train/val/test synthetic JSON/JSONL artifacts
  lora_finetuned_model/     # LoRA adapter + tokenizer assets (LFS tracked)
  gguf_model/               # Exported / quantized GGUF model & tokenizer (LFS tracked)
notebooks/                  # Three sequential workflow notebooks

Quick Start (Local)

Clone the repository and ensure Git LFS is installed so large model artifacts pull correctly:

git clone https://github.com/superlinear-ai/pii-removal-edge-deployment.git
cd pii-removal-edge-deployment
git lfs install
git lfs pull

Running the Notebooks on Google Colab

Prerequisites

• Install Git LFS locally if cloning first.

Steps

1. Clone (or download) the repo locally (optional if you upload files manually):

   git clone https://github.com/superlinear-ai/pii-removal-edge-deployment.git
   cd pii-removal-edge-deployment
   git lfs install
   git lfs pull

2. Open Google Colab.
3. Upload the desired notebook from notebooks/ (File -> Upload notebook).
4. If the notebook requires the model artifacts, either:
   • Upload the needed subfolders from data/ manually, or
   • Add a cell to git clone the repo (Colab runtime) and run git lfs install && git lfs pull.

Git LFS Usage

Large model & adapter artifacts (*.bin, *.pt, *.safetensors, *.gguf) and dataset JSONL files are tracked via patterns in .gitattributes.

Key commands:

git lfs install          # One-time per machine
git lfs track "*.gguf"   # Example: track new pattern
git add .gitattributes
git add path/to/large_file.gguf
git commit -m "Add new quantized model"

To verify LFS pointers:

git lfs ls-files

Adding New Data or Models

1. Place raw large files under an appropriate subfolder in data/.
2. Ensure pattern is in .gitattributes (edit if necessary).
3. Commit via LFS (see above).

Contributing

PRs welcome for improvements, tooling, reproducibility, and inference examples.

License

See LICENSE file.