README.md

---
language:
- en
base_model:
- HuggingFaceTB/SmolLM2-360M-Instruct
---
# LuminAi
![cover image](https://huggingface.co/alvinwongster/LuminAI/resolve/main/asset/LuminAI_logo.png)

## Model Description
**Lumin.AI** is a supportive AI assistant designed to provide immediate emotional support to individuals outside of regular consulting hours. It acts as a supplementary tool for patients and therapists, ensuring that mental health care is more accessible and responsive to users' needs

## Model Demo
![Demo](https://huggingface.co/alvinwongster/LuminAI/resolve/main/asset/demo_2.gif)
![Demo](https://huggingface.co/alvinwongster/LuminAI/resolve/main/asset/demo_1.gif)

## Model Dataset
The chatbot has been trained using [conversational data](https://huggingface.co/alvinwongster/LuminAI/tree/main/dataset), which is supposed to mimick the patient and the therapist. 5 topics where chosen, and 100 conversations from each of these topics were gathered:
- General 
- Relationships 
- Insecurities
- Victim Mentality 
- Self-Improvement 

## How to use
```bash
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("alvinwongster/LuminAI")
model = AutoModelForCausalLM.from_pretrained("alvinwongster/LuminAI")

prompt = "What is depression?"
full_prompt = f"User: {prompt}\nBot:"

inputs = tokenizer(full_prompt, return_tensors="pt")
inputs = {key: val.to(device) for key, val in inputs.items()}

outputs = model.generate(
  **inputs, 
  max_new_tokens=650,
  repetition_penalty=1.3,
  no_repeat_ngram_size=3,  
  temperature=0.8,  
  top_p=0.9, 
  top_k=50  
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)

if "Bot:" in response:
  response = response.split("Bot:")[-1].strip()

print(response)
```

## Model Metrics
To evaluate the chatbot's performance based on our use case, the following weighted metrics system was used:
- Empathy Score (40%): 
    - Measures how well the chatbot responds with empathy.
- Human-Likeness Score (20%):
    - Assesses how natural and human-like the responses feel.
- BERTScore (30%):  
    - Evaluates semantic similarity between chatbot replies and therapist responses. Split equally between F1, Recall and Precision
- Time taken (10%)
    - Time taken to generate a response, a shorter time optimizes user experience 

|Metrics             |GPT  |Llama|LuminAI|
|--------------------|:---:|:---:|:-----:|
|Empathy Score       |0.8  |0.79 |0.79   |
|Human Likeness      |0.27 |0.45 |0.5    |
|BERTScore F1        |0.45 |0.48 |0.51   |
|BERTScore Recall    |0.51 |0.53 |0.55   |
|BERTScore Precision |0.41 |0.44 |0.47   |
|Time Taken          |89.65|15.85|39.42  |
|Total Score         |0.54 |0.65 |0.63   |

## Github Link
Visit [here](https://github.com/alvinnnnnnnnnn/MentalHealth-LLM) for more information on how I trained the model

Try the product [here](https://luminai-chatbot.streamlit.app/)!