Skip to content
GitHub

Training Gym

Distributed training on Modal has never been easier. Pick a framework, plug in a model, and let it jam on your dataset.

Each framework package exposes a factory that returns a modal.App with download_model, prepare_dataset, and train functions.

FrameworkGood forExample
slimeGRPO / RL post-training — Ray + Megatron + SGLangslime_gsm8k, slime_haiku
ms_swiftms-swift Megatron SFT (single- and multi-node)ms_swift_glm_4_7_gsm8k, ms_swift_custom_hf
harborSandbox-based RL with Harbor tasks/agentsharbor_code_golf
pip install git+https://github.com/modal-projects/training-gym.git