Deep Learning Project (P12-A)

This README provides a concise guide to understanding and reproducing the steps involved in the deep learning project focused on fine-tuning the medllama2 model for performing differential diagnosis, utilizing the DDXPlus dataset. Below are the outlined steps and file structures:

A. Dataset folder

1. Data Description: The DDXPlus dataset is stored in the ddxplus-dataset folder inside the parent Dataset directory. It consists of the following files:

   • release_condition.json
   • release_evidences.json
   • test.csv

2. Conversion to Human-Readable Text: The dataset was converted into a more human-readable format stored in cases.json using ddx_data_to_cases.py. This 'cases.json' file consists of 20k samples taken from the test.csv file of DDXPlus.

3. Conversational Data Generation: 10k samples from the cases.json data were further transformed into conversational form between a human and an assistant acting as a doctor using generate.py. The resulting conversational data was stored in generated_data.csv and subsequently pushed to the Hugging Face repository satyam-03/ddx-conversations-10k.

B. Fine-tuning folder

1. Fine-Tuning Script: The fine-tuning process was executed using opdx-alpha.py. This script is responsible for fine-tuning the medllama2 model over our generated DDX conversational dataset, resulting in the output stored in the Model folder.

C. Model folder

1. Model Conversion: The convert-lora-to-ggml.py script in the llama.cpp repository, along with adapter_config.json and adapter_model.bin, were utilized to convert the fine-tuned model into the ggml format and generate the model which can be run in inference mode.

For detailed reference on the above tasks, you can refer to this Medium article.

Running the Model

To run the fine-tuned model, follow these steps:

1. Install Ollama: Ensure that Ollama is installed on your device.

   curl -fsSL https://ollama.com/install.sh | sh
   

2. Start the Ollama server: Run the following command in your terminal:

   ollama serve
   

3. Build the model from fine-tuned adapter: In a new terminal, navigate to the Model directory and run the following commands:

   ollama create opdx -f ./Modelfile
   

   Wait for the process to finish, then run:

   ollama run opdx
   

This command will initiate the model, allowing you to perform differential diagnosis using our fine-tuned opdx model in the terminal itself.

Running the UI

In the UI folder:

1. app folder is the backend of the project.
2. stlit folder contains the frontend made using Streamlit.

To run the and use the UI of project run the run.sh bash script

Note: if run.sh didn't run then:

1. Install the requirements.txt in UI folder
2. Navigate to the UI/app directory and run uvicorn server:app to start the server.
3. In a new terminal, navigate to the UI/stlit/pages directory and then run streamlit run ./Home.py

Final Project Presentation

You can view the final project presentation here.