Medical Advisor Chatbot

⚠️ Project in Progress
>> not yet completed

This project involves fine-tuning the Llama2 model from Huggingface on the lavita/ChatDoctor-HealthCareMagic-100k dataset.

The dataset contains 100k healthcare-related question-answer pairs. It includes real-world user questions posed to medical professionals and their corresponding answers. You can find the dataset here

🔧 The Fine-tuned Model:

Since I was using Colab's GPU, I needed to be mindful of resource usage. To address this challenge, I implemented the following methods in the project:

• 4-bit Quantization: This method reduces the model's memory footprint while maintaining both inference speed and accuracy by using 4-bit quantization.
• LoRA (Low-Rank Adaptation): LoRA is applied to fine-tune only specific layers of the model, such as projection layers, while keeping the core parameters frozen. This approach allows efficient fine-tuning with a smaller number of trainable parameters.

These settings enable the model to perform well in low-resource environments while maintaining the quality of its responses to medical questions.

🔄 Flow of the Agent

Explanation of the Flow:

1. Start: The conversation begins when the user submits a question.
2. ask_user: The system prompts users to provide their questions or input.
3. check_medical: The system analyzes the input to determine if the question is healthcare-related using Word2Vec similarity comparison.
   • If the query is medical, the flow moves to the ask_doctor node.
   • If not, the system checks for greeting intent.
4. check_greeting: If the system detects a greeting from the user, it moves to the ans_greeting node.
   • If the input is neither medical nor a greeting, it directs the user to a fallback response (static_idk_msg), where the system acknowledges that it doesn't understand the input.
5. ask_doctor / ans_greeting / static_idk_msg: Based on the user’s input, the system provides a response:
   • ask_doctor: For healthcare-related queries, the system gives medical advice.
   • ans_greeting: For greetings, the system replies with a greeting.
   • static_idk_msg: If the input is unclear, the system provides a message stating it doesn't understand the question.
6. End: Once the appropriate response is delivered, the conversation ends.