biogpt-large-medicine-instruction-lora

BioGPT-Large Medicine Leaflet Instruction LoRA

This repository contains a LoRA adapter instruction-tuned from microsoft/BioGPT-Large using medicine patient-information leaflet data.

This is a PEFT adapter, not a standalone full model. It must be loaded together with microsoft/BioGPT-Large.

Training

• Method: QLoRA
• LoRA rank: 16
• LoRA alpha: 32
• Target modules: q_proj, v_proj
• Epochs: 3
• Learning rate: 0.0002
• Training examples: 40
• Evaluation examples: 10
• Final evaluation loss: 2.5893

The complete instruction dataset contained 50 examples created from 10 cleaned medicine patient-information leaflets.

Required prompt format

The adapter was trained with the following format:

Instruction:
<question>

Response:

Example:

Instruction:
What is Paracetamol Tablets used for according to the leaflet?

Response:

Plain questions without this wrapper may produce unreliable results.

Loading the adapter

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "microsoft/BioGPT-Large"
adapter_id = "YOUR_USERNAME/biogpt-large-medicine-instruction-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    dtype=torch.float32,
)

model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

instruction = (
    "What is Paracetamol Tablets used for according to the leaflet?"
)
prompt = f"Instruction:\n{instruction}\n\nResponse:\n"

inputs = tokenizer(prompt, return_tensors="pt")

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=100,
        do_sample=False,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

new_tokens = output_ids[0, inputs["input_ids"].shape[1]:]
answer = tokenizer.decode(new_tokens, skip_special_tokens=True)

print(answer)

Use float16 and move the model to CUDA when running on a compatible GPU.

Intended use

This adapter is an educational assignment demonstrating instruction tuning on medicine leaflet content.

Limitations

• It was trained on only 40 examples.
• It is narrowly tuned for medicine-leaflet questions.
• Responses may be incomplete, repetitive, incorrect, or malformed.
• It is not a general-purpose chat model.
• It must not be used for diagnosis, prescribing, dosage decisions, emergency decisions, or other clinical decisions.
• Generated medical information must be verified against an authoritative source and reviewed by a qualified healthcare professional.

Base model

• Base model: microsoft/BioGPT-Large
• Base-model license: MIT