Training Gym
Distributed training on Modal has never been easier. Pick a framework, plug in a model, and let it jam on your dataset.
Tutorials
Runnable notebooks and scripts covering RL post-training, SFT, and infrastructure patterns.
API Reference
Configuration classes, model definitions, and framework launchers — quick lookup for every public API.
Support
Contact points for questions about Training Gym and access to Modal multi-node training.
Frameworks
Section titled “Frameworks”Each framework package exposes a factory that returns a modal.App with download_model, prepare_dataset, and train functions.
| Framework | Good for | Example |
|---|---|---|
slime | GRPO / RL post-training — Ray + Megatron + SGLang | slime_gsm8k, slime_haiku |
ms_swift | ms-swift Megatron SFT (single- and multi-node) | ms_swift_glm_4_7_gsm8k, ms_swift_custom_hf |
harbor | Sandbox-based RL with Harbor tasks/agents | harbor_code_golf |
Install
Section titled “Install”pip install git+https://github.com/modal-projects/training-gym.git