API Reference

API Reference

Complete reference for the training-gym Python library.

Core

Class	Description
ModelConfiguration	Base class for model identity and weight-download logic.
HFModelConfiguration	ModelConfiguration for models hosted on HuggingFace.
ModelArchitecture	Transformer architecture parameters for a specific model.
DatasetConfig	Dataset configuration shared across training frameworks.
WandbConfig	Weights & Biases logging configuration shared across all frameworks.
ModalRayCluster	Base class for bootstrapping a Ray cluster inside Modal clustered functions.
LlmJudge	LLM-as-judge client for an OpenAI-compatible chat-completions endpoint.
TrainResult	One completed training run’s checkpoint handle.

Models

Class	Description
Qwen3-4B	Qwen3-4B (4 billion parameters) from Alibaba.
Qwen3-32B	Qwen3-32B (32 billion parameters) from Alibaba.
GLM-4.7	GLM-4.7 large MoE model from Zhipu AI.
Llama2-7B	Llama 2 7B from Meta.
Kimi-K2.5	Kimi K2.5 from Moonshot AI.

Frameworks

Class	Description
SlimeConfig	slime GRPO training configuration.
ModalConfig (slime)	Modal infrastructure configuration for slime — image setup and dev overlays.
MsSwiftFrameworkConfig	ms-swift Megatron SFT configuration, including Modal infrastructure.
MsSwiftConfig	Top-level wrapper that composes an ms-swift Megatron SFT run.
MilesFrameworkConfig	Miles RLVR configuration, including Modal infrastructure.
MilesConfig	Top-level wrapper that composes a Miles RLVR training run.
HarborFrameworkConfig	Harbor + Miles configuration for sandbox-based RL training.
HarborConfig	Top-level wrapper that composes a Harbor + Miles RLVR training run.