This is large language model (LLM) based on the Zephyr 7B β model. Using the Retrieval Augmented Generation (RAG), the model has been customized to provide answers to questions using information from research papers from the open-source repository arXiv. For over 100K documents, I generated embeddings using HuggingFace's sentence transformer model all-MiniLM-L6-v2 and stored them in a MongoDB database. For each user query, the top 10 documents are found using Atlas Vector Search and used as context for the SciQALM model.
To run this model, get an API token from HuggingFace. Create a file in the root directory named api_token.txt and paste your API token.
Next, create a MongoDB Database called arxiv and a collection named embedded_abstracts. Paste your database connection url in mongodburl.txt. Create a vector search mechanism in MongoDB mechanism based on the abstract_encoded field.
Install the necessary packages by running pip install -r requirements.txt
Run process_arxiv_dataset.py to generate embeddings for the documents in the dataset. Then, run gradio_deploy.py to test the model.
Using Gradio, a user interface to the model was created. Below is an example of a specific question asked and the answer given by SciQALM. It cites specific information from a scientific paper.
A similar model was created using RAG to provide answers to questions related to movies. The code for that model is available in /movie_recommendation and can be run in the same way as SciQALM.