IOI Problem Evaluation

March 18, 2025 ยท View on GitHub

This repository contains code for evaluating Language Models on IOI 2024 problems using LiteLLM.

Installation

  1. Clone the repository
  2. Create a virtual environment with uv (to install uv, follow the UV Installation Guide):
uv venv ioi --python 3.11 && source ioi/bin/activate && uv pip install --upgrade pip
  1. Install dependencies:

uv pip install torch~=2.5.1 --index-url https://download.pytorch.org/whl/cu124
uv pip install sgl-kernel --force-reinstall --no-deps
uv pip install "sglang[all]>=0.4.2.post4" --find-links https://flashinfer.ai/whl/cu124/torch2.5/flashinfer/
uv pip install -r requirements.txt

Environment Setup (In case you want to use remote models)

  1. Copy the environment template:
cp .env.template .env
  1. Edit .env and:
    • Uncomment the variables for the LLM providers you plan to use
    • Replace the placeholder values with your actual API keys
    • Optional: Configure proxy settings if needed

Example .env for using OpenAI's GPT-4:

OPENAI_API_KEY=your_actual_key_here
OPENAI_ORGANIZATION=your_org_id  # Optional

Usage

Running with Remote Models

Run the evaluation with remote models:

python evaluate.py --org_id YOUR_ORG_ID --model_id YOUR_MODEL_ID [--num_generations 50] [--concurrency 5]

Command line arguments:

  • --org_id: Organization ID (required)
  • --model_id: Model ID in LiteLLM format (required)
  • --api_base: API base URL for the model (optional)
  • --num_generations: Number of generations per problem (default: 50)
  • --num_retries: Number of retries for failed API calls (default: 10)
  • --concurrency: Number of concurrent generations (default: 20)
  • --num_problems: Number of problems to evaluate (default: all)
  • --num_subtasks: Number of subtasks to evaluate per problem (default: 1, use -1 for all)
  • --dry_run: Run without making actual LLM calls
  • --override: Override existing results and start fresh
  • --model_postfix: Postfix for the model name
  • --revision: Revision to use for the model
  • --timeout: Timeout for the LLM call in seconds (default: 600)
  • --use_requests: Use requests instead of litellm
  • --max_tokens: Maximum number of tokens for generation

Running with Locally Deployed Models (SGLang)

For locally deployed models using SGLang, you can use the provided scripts:

Using SLURM for Distributed Deployment

For HPC environments with SLURM, use run_ioi_slurm.py to evaluate open models:

python run_ioi_slurm.py --model "MODEL_PATH" --concurrency 30 --startup_delay 7200 --logs_dir "DIR_FOR_OUTPUT_LOGS" --slurm_dir "DIR_FOR_SLUR_SCRIPT" --uv_env "PATH_TO_UV_ENV" --eval_args "--org_id YOUR_ORG_ID"

Output

The results will be saved in directory specified by --logs_dir with structure:

{org_id}/{revision}-{model_id}-{postfix}/

The output includes:

  • Generated code solutions for each problem and subtask
  • Metrics on generation performance
  • Token usage statistics

You can analyze the results using the saved data to evaluate the model's performance on competitive programming tasks.