HealthApp

Project Overview

A fine-tuned Large Language Model (LLM) designed to identify user emotional states and provide empathetic support.

Key Features

• Emotion Detection: Accurately identifies user emotional states, including happiness, sadness, anger, and fear.
• Emotional Support: Provides empathetic and supportive responses to users.
• Conversational Interface: Engages in natural-sounding conversations with users.

Technical Specifications

• Base Model: LLaMA 3.2 (3B)
• Dataset: Mental Health Counseling Conversations
• Fine-tuning Framework: Unsloth
• Fine-tuning Method: LoRA (Low-Rank Adaptation) for efficient adaptation
• Parameter Quantization: 4-bit parameter quantization for reduced computational requirements

Usage

Installation

1. Install the required libraries:

   pip install unsloth transformers datasets torch

2. Log in to Hugging Face to access the model:

   from huggingface_hub import login
   from kaggle_secrets import UserSecretsClient
   
   user_secrets = UserSecretsClient()
   hf_token = user_secrets.get_secret("HUGGINGFACE_TOKEN")
   login(hf_token)

3. Download the pre-trained model and tokenizer:

   from unsloth import FastLanguageModel
   
   model, tokenizer = FastLanguageModel.from_pretrained(
       model_name="unsloth/Llama-3.2-3B-Instruct",
       max_seq_length=2048,
       dtype=None,
       load_in_4bit=True,
   )

Example Usage

from unsloth import FastLanguageModel

def get_response(user_input):
    instruction = """You are an experienced psychologist named Amelia.\nBe polite to the patient and answer all mental health-related queries."""

    messages = [
        {"role": "system", "content": instruction},
        {"role": "user", "content": user_input}
    ]

    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(prompt, return_tensors='pt', padding=True, truncation=True).to("cuda")
    outputs = model.generate(**inputs, max_new_tokens=150, num_return_sequences=1)

    text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return text.split("assistant")[1]

# Test the model
user_input = "I'm feeling really sad today."
response = get_response(user_input)
print(response)

License

Our Emotional Support LLM is licensed under [Insert License].

Acknowledgments

We thank the creators of LLaMA, Unsloth, and LoRA for their contributions to our project.

Future Development

• Multilingual Support: Expand language support for global accessibility.
• Emotion Analysis: Integrate advanced emotion analysis for more accurate support.
• Human Evaluation: Conduct human evaluations for model refinement.