ACM SGU Competitive Programming Solutions with LLM Enhancement

Overview

This repository contains solutions to ACM SGU (Saratov State University) competitive programming problems, enhanced with detailed editorials and reasoning explanations generated using advanced language models, and includes a finetuned model trained specifically on this dataset.

The repository consists of three main components:

• Original Solutions: Competitive programming solutions to SGU problems in C++ or Python.
• Enhanced Editorials: Comprehensive problem explanations, solution approaches, and detailed editorials generated using o4-mini.
• Finetuned Model: A basic pipeline with MLX for finetuning an LLM based on the above (still WIP).

Dataset Structure

dataset/
├── p100.txt          # Enhanced editorial with solution approach
├── p100_raw.txt      # Original source code solution + statement + sample input/output
├── p101.txt          # Enhanced editorial with solution approach  
├── p101_raw.txt      # Original source code solution + statement + sample input/output
└── ...               # Additional problems (100+ problems covered)

Enhanced Editorial Format

Each enhanced editorial (p*.txt) contains:

• Concise problem statement
• Detailed solution approach and algorithm explanation
• Step-by-step implementation guide
• Time/space complexity analysis
• Alternative solution methods
• C++ and Python reference implementations

Model Finetuning

Dataset: 181 problems (and slowly increasing) from the SGU archive solved by me, enhanced with detailed editorials using GPT o4-mini.

Training Configuration:

• Base Model: OlympicCoder-7B (MLX format, fp16)
• Architecture: Qwen2-based language model
• Training Method: Autoregressive language modeling with SGD optimizer
• Context Length: 2048 tokens with sliding window
• Learning Rate: 1e-6 with gradient clipping
• Hardware: Apple M2 Max (96GB RAM)

Getting Started

# Install MLX for Apple Silicon
pip install mlx mlx-lm

# Install additional dependencies  
pip install matplotlib numpy transformers

Training Your Own Model

python3 src/finetune.py \
    --model path/to/base/model \
    --dataset_dir dataset \
    --epochs 3 \
    --batch_size 1 \
    --lr 1e-6 \
    --max_length 2048 \
    --output_dir path/to/output