Shiva7706/RxStruct-Gemma-1B

Fine-tuned Model: RxStruct-Gemma-1B | Quantized Version: GGUF Release

# RxStruct-Gemma-1B (LoRA)

A fine-tuned variant of Gemma-3-1B-IT optimized for structured medical data extraction from natural, doctor-style prescription dialogues.
This model outputs fully structured JSON containing medicines, dosages, diseases, tests, and instructions — without requiring external post-processing.

---

Model Overview

Property	Value
Base Model	google/gemma-3-1b-it
Fine-tuning Framework	Unsloth
Method	LoRA (Rank=8, α=16, Dropout=0.05)
Precision	bfloat16
Sequence Length	1024 tokens
Stop Token	"AAA"
Parameters Trained	~13M (1.29%)
Dataset Source	Synthetic Claude 3.5 Sonnet-generated doctor–patient prescription conversations
Output Format	Valid JSON object with fixed schema

---

Example Usage

from unsloth import FastLanguageModel
from transformers import TextStreamer

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="google/gemma-3-1b-it",
    adapter_name="Shiva7706/RxStruct-Gemma-1B",
)
FastLanguageModel.for_inference(model)

prompt = """Mr. Shah, your blood pressure is quite high at 160/100.
I'm starting you on Amlodipine 5mg once daily in the morning.
Also take Atorvastatin 10mg at bedtime for your cholesterol.
Get your lipid profile and kidney function tests done after 1 month.
Reduce salt intake and exercise regularly."""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer=streamer, max_new_tokens=512)

## Example Output
```json
{
  "medicines": [
    {"name": "Amlodipine", "dosage": "5mg", "frequency": "once daily", "duration": "unspecified", "route": "oral", "timing": "morning"},
    {"name": "Atorvastatin", "dosage": "10mg", "frequency": "at bedtime", "duration": "unspecified", "route": "oral", "timing": "unspecified"}
  ],
  "diseases": ["high blood pressure"],
  "symptoms": ["high blood pressure"],
  "tests": [
    {"name": "lipid profile", "timing": "after 1 month"},
    {"name": "kidney function tests", "timing": "after 1 month"}
  ],
  "instructions": ["reduce salt intake", "exercise regularly"]
}

Training Details

Component	Specification
GPU	NVIDIA GeForce RTX 3050 Laptop GPU (6 GB VRAM)
Peak VRAM Usage	~2.52 GB
System RAM	16 GB (Dell G15)
CUDA Version	13.0
Driver Version	581.08
Frameworks	PyTorch 2.8.0 + CUDA 12.8 + Unsloth 2025.10.7
Training Time	~7 minutes / 3 epochs / 166 samples
Validation Loss	0.2435
Validation Perplexity	1.28

Training was performed in a Linux WSL2 Devcontainer environment with gradient offloading and memory optimization enabled through Unsloth.

Dataset

The dataset consists of synthetic, Indian-style prescription conversations generated with Claude 3.5 Sonnet, following strict extraction rules.

Schema:

{
  "medicines": [
    {"name": "string", "dosage": "string", "frequency": "string", "duration": "string", "route": "string", "timing": "string"}
  ],
  "diseases": ["string"],
  "symptoms": ["string"],
  "tests": [{"name": "string", "timing": "string"}],
  "instructions": ["string"]
}

All conversations are synthetic and do not contain any personally identifiable or real patient data.

Model Performance

• Validation Loss: 0.2435
• Validation Perplexity: 1.28
• JSON Structural Accuracy: ~94% (measured on 50 random generations)
• Inference Latency (RTX 3050): ~1.9s per 300-token generation

Limitations

• The model is trained only on synthetic data, not real medical transcripts.
• It should not be used for clinical decision-making.
• Certain ambiguous dialogues may lead to redundant entities (e.g., mixing tests and medicines).
• JSON format adherence is strong but not perfect; a small post-processor is recommended.

Recommended Post-Processing (Optional)

import json, re

def clean_json_output(text):
    match = re.search(r"\{[\s\S]*\}", text)
    if match:
        text = match.group(0)
        try:
            data = json.loads(text)
            if "AAA" in text:
                text = text.replace("AAA", "")
            return data
        except json.JSONDecodeError:
            return text[:text.rfind("}")+1]
    return None

Intended Use

Intended for:

• Research on medical NLP and structured data extraction
• Building medical assistants that convert prescriptions to structured EHR-compatible data
• Educational and demonstration purposes

Not intended for:

• Real-world clinical applications
• Diagnostic or treatment decision systems