CP/DSA Question Solver

Team Name: ML 35

Welcome to the CP/DSA Question Solver project, part of the IITI Summer of Code initiative! This repository contains the code and resources for building an automated solver for competitive programming and data structures & algorithms (CP/DSA) questions from various platforms.

Overview

This project aims to create an automated solver for CP/DSA questions using state-of-the-art deep learning techniques. The solver is divided into three main workflows, each tailored to different platforms: Leetcode, Codeforces, and others.

Common Workflow

1. Question Simplification:

   • Tool Used: Phi-3
   • Simplifies complex questions to make them more manageable for the model.

2. Retrieval-Augmented Generation (RAG):

   • Retrieval: Uses SentenceTransformer MiniLM and FAISS to retrieve relevant information.
   • Model: Specific models are used depending on the platform (detailed below).

3. Fine-Tuning LLM:

   • Dataset: Consists of 'prompt','accepted','rejected' columns for ORPO fine-tuning , made using base LLM and pandas.
   • Method: Reinforcement Learning based ORPO fine tuning, straying away from 'rejected' answers and running parallel to 'accepted' answers.

4. Scraping Questions:

   • Tools Used: BeautifulSoup and GitHub API
   • Collects questions from the respective platforms for training and testing the models.

5. Deployment:

   • Tool Used: Streamlit
   • Provides a user-friendly interface to interact with the question solver.

Platform-Specific Workflows

1. Leetcode Workflow

For solving Leetcode questions:

• Model: Implements a RAG approach tailored for Leetcode questions.
• Deployment: Streamlit is used to deploy the Leetcode question solver.

2. Codeforces Workflow

For solving Codeforces questions:

• Model: Uses a DeepSeek-Coder-6.7B-base model, fine-tuned with RL-based ORPO techniques.
• Deployment: Streamlit is used to deploy the Codeforces question solver.

3. New Questions Workflow

For solving questions from other CP/DSA platforms:

• Model: Integrates RAG with DeepSeek-V2-Lite-Instruct (16B version) for a versatile solution.
• Deployment: Streamlit is used to deploy the solver for other platforms.

Technologies Used

• Deep Learning Models:

  • DeepSeek-Coder-6.7B-base
  • DeepSeek-V2-Lite-Instruct (16B version)
  • Phi-3 for question simplification

• Libraries and Tools:

  • RAG Implementation: SentenceTransformer MiniLM, FAISS
  • Scraping: BeautifulSoup, GitHub API
  • Deployment: Streamlit

Interface:

Team Members:

1)Chanakya Cherukumalli
2)Varad Pendse
3)Rishab Anand
4)Kushagra Mishra