14B_GRPO_matharena

<div align="center"> <h1><img height="150px" src="./images/matharena_icon.png" alt="MathArena"><br>MathArena</h1> <a href="https://www.python.org/"> <img alt="Build" src="https://img.shields.io/badge/Python-3.12-1f425f.svg?color=blue"> </a> <a href="https://opensource.org/licenses/MIT"> <img alt="License: MIT" src="https://img.shields.io/badge/License-MIT-yellow.svg"> </a> </div>

👋 Overview

MathArena is a platform for the evaluation of LLMs on the latest math competitions and olympiads. It is hosted on matharena.ai. This repository contains all the code used for model evaluation of the competitions. The README explains how to run your models or add a new competition.

📑 Table of Contents

• Installation
• Evaluating a New Model
  • Model Configuration
  • Running the Model
  • Local VLLM Usage
• Adding a Competition
  • Setting Up Competition Files
  • Verifying Problem Statements
  • Running Models on Competitions
  • Competitions Requiring Grading
  • Running LLMs as Judges
• Viewing Results
• Evaluation logs

🚀 Installation

MathArena uses UV to manage dependencies. If you want to run local models, uncomment the vllm installation in pyproject.toml.

Install UV

• macOS and Linux:

  curl -LsSf https://astral.sh/uv/install.sh | sh
  

• Windows:

  powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
  

---

Alternative installation

As an alternative to UV, you can also create a conda environment and install the package as follows:

conda create -n matharena python=3.12
conda activate matharena
python -m pip install -e .

If you choose this option, disregard uv run in all instructions and use python directly instead.

🤖 Evaluating a New Model

Model Configuration

Create a configuration file in the configs/ folder. Each config must include:

• Required:
  • model: Model name. Reasoning effort of OpenAI models can be set by appending --[low/medium/high] to the model name, e.g., o3-mini--high.
  • api: API provider. The API key should be defined as environment variable when using the specified API. The supported options with their corresponding API keys are:
    • openai: OPENAI_API_KEY
    • anthropic: ANTHROPIC_API_KEY
    • together: TOGETHER_API_KEY
    • google: GOOGLE_API_KEY
    • deepseek: DEEPSEEK_API_KEY
    • openrouter: OPENROUTER_API_KEY
    • vllm: (runs locally; no API key required)
  • human_readable_id: A unique, descriptive identifier.
• Optional Parameters:
  • API settings like temperature, top_p, and top_k (default: temperature is from competition config, see Adding a Competition).
  • max_tokens: Max number of tokens for the model (default: from competition config, see Adding a Competition).
  • concurrent_requests: Number of parallel requests to API (default: 30).
  • timeout: Request timeout in seconds (default: 2000).
  • max_retries: Retry attempts to API (default: 50).
  • read_cost & write_cost: Cost per million tokens in USD for input and output tokens (default: 1 each).
  • date: Creation date of the model in the format "yyyy-mm-dd". Only affects the website.
  • batch_processing: If set to true, the model will be queried using batch processing. Only available for OpenAI and Anthropic models.

Running the model

Execute the following command to evaluate a model on a competition:

uv run python scripts/run.py --configs path/to/your/config --comp path/to/competition

• path/to/your/config: Relative path from the configs/ folder to the model configuration.
• path/to/competition: Relative path from the data/ folder to the competition folder.

Example:

uv run python scripts/run.py --configs openai/gpt-4o.yaml --comp aime/aime_2025_I

Additional Flags:

• skip_existing: Skip problems already processed through the model.
• n: Number of runs per problem (default: 4).

Note: Errors thrown by the API provider are retried every minute up to 50 times. If no answer is returned after 50 tries, the answer will be counted as incorrect. Running again with skip_existing enabled will attempt to run the problems on which this occurred again.

Running Models Locally Using VLLM

If using a local model with vllm, start the server:

vllm serve [[model_name]] --dtype auto --api-key token-abc123

➕ Adding a Competition

Adding a competition can be done in several quick steps.

Setting Up Competition Files

In the data/ folder, create a new directory for your competition with the following structure:

1. Problems:
   • Create a subfolder problems/ and add each problem as a separate LaTeX file named 1.tex, 2.tex, ..., {k}.tex, where k is the number of problems in your competition. You can skip a problem if you want/need to.
2. Competition Config:
   • Create a config.yaml with:
     • instruction: Instructions for the model. Must require the final answer be in \boxed{}.
     • default_temperature: Default temperature.
     • default_max_tokens: Default max tokens.
     • strict_parsing: true for strict format matching (e.g., only \boxed{43} is accepted) or false for lenient parsing.
     • n_problems: Total number of problems.
     • date: Date of the competition, in the format "YYYY-MM-DD".
     • final_answer (optional): If set to false, the competition is one that is manually graded with judges. Defaults to true if not set.
3. Answers:
   • If the competition is one based on final answers, add an answers.csv file with columns id and answer.
     • id: The problem filename (without the .tex extension).
     • answer: The integer answer.
   • If the competition is evaluated using human judges, add a grading_scheme.json file. This file should consist of a list of dictionaries, each of which contain the following fields:
     • id: The problem filename (without the .tex extension).
     • points: The maximum number of points for the question.
     • scheme: A list of dictionaries, each containing substeps for which points are awarded. Each dictionary contains the following keys:
       • points: Points associated with this step.
       • title: Title of the step. Should be unique across all dictionaries in this scheme.
       • desc: Description of the step.

Verifying Problem Statements

Ensure your LaTeX problems compile correctly:

uv run python scripts/check_latex.py --comp path/to/competition

Then, build the latex/main.tex to generate a PDF and confirm all problems appear as expected.

Running Models on Competitions

To run multiple models (possibly across different APIs), use:

uv run python scripts/run_multiple.py --apis openai google anthropic together --comp path/to/competition

This will run models from the same API sequentially and from different APIs concurrently. Options:

• --simul: Run all models in parallel, even if they use the same API.
• models: Provide space-separated regex patterns to filter models. A model is only run if it matches any of the regexes.
• skip_existing: Skip problems already processed through the model.
• n: Number of runs per problem (default: 4).

Note: For local vllm usage, ensure the vllm server is running as described above. Logs will be found in the logs/ folder.

Competitions Requiring Grading

To set up grading of questions, convert the model answers to TeX files:

uv run python scripts/judge/answers_to_latex.py --comp path/to/competition

This will compile all model answers in a PDF file in the folder latex/path/to/competition.

Now, collect all PDFs for all evaluated models in a single folder using:

uv run python scripts/judge/collect_pdfs.py --comp path/to/competition

This will put all PDFs associated with question with idx i in the folder latex/path/to/competition/i. Each PDF will be given a unique (anonymous) ID. Follow the instructions in README_judges.md to grade each PDF. The grading of a single PDF should be placed in grading/path/to/competition/i/{ID}.json, where the ID is the ID given to the PDF associated with the grading. In case a PDF is graded by multiple people, you can add more files by naming them grading/path/to/competition/i/{ID}_{X}.json where X is any suffix. Finally, run

uv run python scripts/judge/merge_judgments.py --comp path/to/competition

This will add the judgments directly in the raw output traces of the models.

Running LLMs as judges

To run an LLM as a judge, you must first add the solutions of all problems of the competition in data/path/to/competition/solutions/{i}.text where i is the index of the problem.

Then, use the following command:

uv run python scripts/grade.py --grader_config path/to/grader --solver_config path/to/solver/1 path/to/solver/2 --comp path/to/competition

Options:

• path/to/grader: Relative path from the configs/ folder to the model configuration for the judge.
• path/to/solver: Relative path from the configs/ folder to the model configuration of the judged model. Multiple ones can be given by passing space-separated paths.
• path/to/competition: Relative path from the data/ folder to the competition folder.

Example:

uv run python scripts/grade.py --grader_config openai/o3-mini.yaml --solver_config openai/o3-mini.yaml anthropic/claude-37.yaml --comp usamo/usamo_2025

Additional Flags:

• skip_existing: Skip problems already processed through the model.
• n: Number of runs per problem to evaluate (default: 4). Must be no larger than the amount of generated solutions.

Notes: For local vllm usage, ensure the vllm server is running as described above. Logs will be found in the logs/ folder. We also recommend either using generalist models or either of o1, o3-mini or Claude 3.7 as graders due to their robustness with respect to the formatting instructions.

After obtaining the judgments, you can then process them using the aforementioned merge_judgments.py script:

uv run python scripts/judge/merge_judgments.py --comp path/to/competition --grading-folder autogradings

This will add the judgments directly in the raw output traces of the models.

📊 Viewing Results

Launch a local web server to inspect the results:

uv run python scripts/app.py --comp path/to/competition

Access the app at http://localhost:5001/. Warning signs for solutions indicate a potential problem with the model run and should be manually verified. Any warning is caused by one of the following problems:

• 💀: parser threw an error or encountered something unexpected.
• ⚠️: The correct answer might be present in the model answer, but it was not extracted.
• ❕: Model likely hit max token limit.

If issues are found, delete the corresponding output file or fix the parser and rerun the model with skip_existing. If the parser requires a manual overwrite, you can edit src/matharena/parse_manual.py and add a key-value pair mapping the model solution to a parseable solution.

🪵 Evaluation Logs

You can find logs from our evaluation containing full reasoning traces (if available) and solutions produced by the models at the following link: https://files.sri.inf.ethz.ch/matharena/matharena_data.zip.

📚 Citation

@misc{balunovic_srimatharena_2025,
  title = {MathArena: Evaluating LLMs on Uncontaminated Math Competitions},
  author = {Mislav Balunović and Jasper Dekoninck and Ivo Petrov and Nikola Jovanović and Martin Vechev},
  copyright = {MIT},
  url = {https://matharena.ai/},
  publisher = {SRI Lab, ETH Zurich},
  month = feb,
  year = {2025},
}