harmony-clinical-structuring-qwen25-3b-lora-200step

Harmony Clinical Structuring QLoRA Adapter

This repository contains a proof-of-concept QLoRA adapter for structured clinical information extraction.

Base model

• Qwen/Qwen2.5-3B-Instruct

Task

The model takes clinical text as input and returns schema-valid JSON containing extracted:

• medications
• adverse events
• dosage attributes where available
• linked medication information
• evidence text
• source spans

Training setup

• Method: QLoRA
• Quantization: 4-bit NF4
• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• Training steps: 200
• Dataset: ADE Corpus v2
• Frameworks: Hugging Face Transformers, PEFT, TRL, PyTorch

Intended project use

In the Harmony Healthcare RAG workflow, this adapter is intended to run at ingestion time after OCR and chunking. It enriches text chunks with structured medication and adverse-event metadata for SQLite storage, search filtering, review, and UI display.

Status

This is a project/demo proof-of-concept adapter, not a production clinical model.

Limitations

• Not clinically validated
• Not for diagnosis or treatment decisions
• Not production-ready
• Longer chunked training is recommended for stronger span accuracy