A clean, modular experiment framework for math reasoning with Hydra configs.

conda env create -f environment.yml
conda activate pita

python run_parallel.py experiment.name=my_experiment

The parallel runner automatically:

• Detects available GPUs
• Parallelizes data generation across GPUs
• Distributes training jobs optimally
• Works seamlessly with 1 GPU or multiple GPUs

Hydra will create an output directory at outputs/<date>/<experiment-name>-<time>/.

If a run fails or is interrupted, you can resume from where it left off:

python run_parallel.py resume_from=outputs/2025-10-20/my_experiment-14-30-00

sh run_experiments.sh