Welcome to MediBuddy! Here we provide basic answers for your medical queries Your First Line Of Care.
- A large language model which acts as a recommendation system which runs locally on your PC
- Can identify vast number of diseases based on given symptoms and be a robust first responder to give precautions as required
- Restricated and ultra specialized for pediatric use
- Based on Meta AI’s state of the art open source AI-basesd LLM model
- Custom fine-tuned model based on medical textbooks including Gale Encyclopedia of Medicine,Essential Pediatric Medicine as well as Medical Pharmology
- LLM(LLAMA 2.0 7B)
- Vector DataBase(FAISS)
- LangChain
- ChainLit
- Data Collection- Acquired various Textbooks of medicine curated by doctors to train our LLM.
- Data Preprocess-PDF's of Textbooks were fed to LangChain Document Ingestor. It takes the PDF and applies advanced OCR tools to take the text data and add it to the document in a recursive manner using 500 tokens for each document and overlapping 50 tokens.
- VectorDB Creation - A vector database is a type of database that stores data as high-dimensional vectors, which are mathematical representations of features or attributes. It mainly stored the embeddings of the data we created. We use FAISS(Facebook AI Similarity Search) created by Facebook to store and search the similarity between the query embedding and the embedding in the vector database.
- Model Inference-Inference is done using LLaMA model which was introduced by the Meta AI Research team and is a collection of many transformers and has 7B parameters.
- Fine Tuning -The model was fine-tuned on new Medical data and restrictions were imposed by setting a pre prompt that sets the conditions and what the output should be the limit of what the model outputs
The weight file for the llama-2-7b is available here. Add this file to the root directory after cloning the repository.
-
This project was a submission for the HackZ hackathon conducted by Osmania Business Technology Incubator and we won the first prize for the problem statement of Pediatric Disease Predictor and First Responder