Skip to content

Latest commit

 

History

History
75 lines (58 loc) · 4.46 KB

README.md

File metadata and controls

75 lines (58 loc) · 4.46 KB

Domain-Specific-LLM-Finetuning-on-Finance-Dataset-

FinancialDataUsingLLM

AI-Driven Question Answering System

Overview

This repository contains the implementation of an AI-driven Question Answering System that leverages state-of-the-art machine learning techniques and NLP models to efficiently retrieve accurate answers from a large corpus of documents. This system integrates FAISS for efficient similarity search in high-dimensional spaces and utilizes models like OpenAI's GPT-3 and fine-tuned transformers for generating responses.

Project Structure

  • FAISS_INDEX.py: Script for managing the FAISS index and saving the vector database.
  • QnA.py: Core script containing the model logic, including functions to generate answers and find relevant document chunks.
  • evaluation.py: Script for generating predictions on a sample set and saving the results for metric calculations.
  • extract_urls.py: Utility for scraping URLs from the Moneycontrol website.
  • faiss_index.pkl: Pickle file containing preprocessed document chunks stored in a vector database format.
  • file_info.json: JSON file containing metadata about the document chunks such as file counts and URLs.
  • finetuning.ipynb: Jupyter notebook detailing the fine-tuning process of the models used in the project.
  • load_documents.py: Script to display the contents of the pickle file containing the document chunks.
  • main.py: Streamlit application script for the user interface, including functionality to display metrics graphs.
  • metrics.json: JSON file containing computed metrics from model evaluations.
  • metrics_evaluation.py: Script that computes metrics based on model predictions and reference answers.
  • record_metrics_data_huggingface.json: Output data containing predictions for sample questions using the Hugging Face model.
  • record_metrics_data_openai.json: Output data containing predictions for sample questions using OpenAI's model.
  • requirements.txt: List of Python libraries required to run the project.
  • sample.json: Sample data file containing questions, reference answers, and contexts used for metric calculations.

Setup and Installation

  1. Clone the Repository: 
    git clone https://github.com/ManojKumarKolli/FinancialDataUsingLLM.git
cd FinancialDataUsingLLM/QuestionAndAnswerUsingLangchainInFinance
  2. Install Dependencies: Make sure Python 3.8+ is installed on your system, then run: 
    pip install -r requirements.txt
  3. Running the Streamlit Application: Execute the following command to start the Streamlit interface: 
    streamlit run main.py

Setting Up API Keys

To ensure the project functions correctly, you'll need to obtain API keys from OpenAI and Weights & Biases. Follow these steps to set up your environment:

OpenAI API Key

  1. Obtain Key:
    • Visit OpenAI and sign up or log in.
    • Navigate to the API section and generate a new API key.

Weights & Biases API Key

  1. Obtain Key:
    • Sign up or log in at Weights & Biases.
    • Access your user settings and navigate to API keys to generate a new key.

Storing API Keys Securely

  1. Create a .env File in your project's root directory.
  2. Wand api key is used in fine-tuning.
  3. Add the API Keys to the .env file: 
    open_api_key=YOUR_OPENAI_API_KEY_HERE
WANDB_API_KEY=YOUR_WANDB_API_KEY_HERE

Contributions

We extend our deepest gratitude to each team member whose dedicated efforts have significantly shaped this project:

  • Manoj Kumar (@manojkumar): Led the integration and optimization of the OpenAI model, managed the model fine-tuning processes, developed evaluation metrics, and contributed to human-centric metrics analysis.
  • Meghana Reddy (@meghanaDasireddy): Responsible for data scraping and preprocessing, managed the FAISS indexing system, and played a key role in the evaluation of human-centric metrics.
  • Krishna Sai (@ananthakrishna): Developed the Streamlit application, integrated the Hugging Face model, implemented metric plotting in the UI, and handled various evaluation processes including human-centric metrics.

Acknowledgments

Special thanks to all mentors and advisors who provided guidance throughout the project. Their insights were invaluable in navigating the complexities of advanced NLP and system integration.