Fine-Tuning GEMMA-2B for Doctor-Patient Interaction: Efficient Model Adaptation Using LoRA and 4-bit Quantization
Model Description
This model is a fine-tuned version of GEMMA-2B, adapted specifically for doctor-patient interaction tasks. Fine-tuning was performed using Low-Rank Adaptation (LoRA) and 4-bit quantization for efficient model adaptation. The model aims to facilitate improved, context-aware, and relevant interactions in medical consultations, enhancing communication and understanding between healthcare providers and patients.
Developed by: Shailesh Kumar Khanchandani
Shared by: Shailesh Kumar Khanchandani
Model type: Causal Language Model
Language(s) (NLP): English
Finetuned from model: google/gemma-2b
Model Sources
Repository: Fine-Tuned GEMMA-2B
Uses
Direct Use
The model is intended for direct use in generating contextually appropriate responses for doctor-patient interactions. It can be used in virtual assistants, chatbots, and other medical consultation platforms to assist healthcare providers in communicating with patients.
Downstream Use
The model can be further fine-tuned for specific medical specialties or customized to fit the unique requirements of different healthcare systems or languages.
Out-of-Scope Use
The model is not intended for use in generating medical advice or diagnoses without human oversight. It should not be used in any scenario where critical health decisions are made solely based on its outputs.
Bias, Risks, and Limitations
The model inherits biases present in the training data. Users should be cautious of potential biases and inaccuracies, particularly in sensitive medical contexts.
Recommendations
Users should ensure that outputs are reviewed by qualified healthcare professionals before being utilized in a clinical setting. Continuous monitoring and updating of the model with diverse and representative data can help mitigate biases.
How to Get Started with the Model
Use the code below to get started with the model:
Python Code
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
def generate_text(prompt):
# Model and tokenizer initialization
model_name = "skkjodhpur/Gemma-Code-Instruct-Finetune-by-skk"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
# Move model to GPU if available
device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)
# Tokenize input
input_ids = tokenizer.encode(f"<s>[INST] {prompt} [/INST]", return_tensors="pt").to(device)
# Generate text
with torch.no_grad():
output = model.generate(
input_ids,
max_length=200,
num_return_sequences=1,
do_sample=True,
temperature=0.7,
)
# Decode and return the generated text
generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
return generated_text
# Example usage
prompt = "I fell on uneven sidewalk on July 3rd and hurt my knee: I had no problem walking and I didn t have much pain just bruising on my knee with tenderness. Now as the bruising is entering the purple/green phase I am starting to get a lot of swelling in my calf and ankle- what should I do ??"
response = generate_text(prompt)
print("Generated response:")
print(response)
Output:
Generated response:
[INST] I fell on uneven sidewalk on July 3rd and hurt my knee: I had no problem walking and I didn t have much pain just bruising on my knee with tenderness. Now as the bruising is entering the purple/green phase I am starting to get a lot of swelling in my calf and ankle- what should I do ?? [/INST]
Output:Take care Chat Doctor - Avoid walking and running to protect your knee .- Take pain relievers to help reduce pain and swelling .- Take cold compresses to help reduce pain and swelling .- Use crutches to assist with walking for a few days to protect your knee.
Training Details
Training Data
The training data includes dialogues from the "Doctor-Chat-Dataset," which consists of anonymized conversations between doctors and patients. The dataset was processed to ensure privacy and relevancy.
Training Procedure
Preprocessing
The input texts were tokenized using AutoTokenizer from the transformers library. The dataset was mapped to the tokenized format and formatted appropriately for causal language modeling.
Training Hyperparameters
Training regime: Mixed precision (fp16)
Batch size: 1
Gradient accumulation steps: 4
Warmup steps: 0.03
Max steps: 100
Learning rate: 2e-4
Optimizer: PagedAdamW (8-bit)
Save strategy: Per epoch
Technical Specifications
Model Architecture and Objective
The model follows the architecture of a causal language model, finetuned for specific tasks using parameter-efficient fine-tuning techniques (LoRA) and quantized to 4-bit precision for efficient deployment.
