API Reference

Complete reference for the training-gym Python library.

Core

Class	Description
ModelConfig	Base class for model identity and weight-download logic.
HFModelConfiguration	ModelConfig for models hosted on HuggingFace.
ModelArchitecture	Transformer architecture parameters for a specific model.
DatasetConfig	Dataset configuration shared across training frameworks.
HuggingFaceDataset	Dataset backed by a HuggingFace datasets repo.
HarborDataset	Dataset backed by a Harbor task directory structure.
WandbConfig	Weights & Biases logging configuration shared across all frameworks.
ModalRayCluster	Base class for bootstrapping a Ray cluster inside Modal clustered functions.
TrainResult	One completed training run’s checkpoint handle.

Evaluation

Class	Description
EvalConfig	Evaluate a deployed model on a dataset config.
EvalResult	Saved results for one evaluation run across a dataset.
EvalRowResult	One model interaction: the prompt, the raw response, its parsed
HarborEval	Evaluate a deployed model on a Harbor dataset using sandbox execution.

Models

Class	Description
ToolCall	A parsed tool invocation from model output.
ParsedResponse	Structured result of parsing raw model output.
parse_qwen3_response	Parse Qwen3-family model output into structured content.
Qwen3-0.6B	Qwen3-0.6B (0.6 billion parameters) from Alibaba.
Qwen3-1.7B	Qwen3-1.7B (1.7 billion parameters) from Alibaba.
Qwen3-4B	Qwen3-4B (4 billion parameters) from Alibaba.
Qwen3-8B	Qwen3-8B (8 billion parameters) from Alibaba.
Qwen3-14B	Qwen3-14B (14 billion parameters) from Alibaba.
Qwen3-30B-A3B	Qwen3-30B-A3B (30B total, ~3B active) MoE model from Alibaba.
Qwen3-32B	Qwen3-32B (32 billion parameters) from Alibaba.
Qwen3.6-27B	Qwen3.6-27B (27B-parameter dense) model from Alibaba.
Qwen3.6-35B-A3B	Qwen3.6-35B-A3B (35B total, ~3B active) MoE model from Alibaba.

Training

Class	Description
TrainConfig	Compose dataset, model, and recipe into one training entrypoint.
MultiTurn	Configure multi-turn rollout for conversational RL training.
SlimeRecipe	Recipe dataclass for configuring slime GRPO training on Modal.
Qwen3_6_27b_Recipe	Qwen3.6-27B dense hybrid model on 1×8×H100 with TP4×PP2, colocated GRPO.
Qwen3_6_35b_Recipe	Qwen3.6-35B-A3B (MoE) on 1×8×H100 with TP2/PP2/CP1/EP4.

Deployment

Class	Description
DeploymentConfig	Deploy a model behind a serving engine.
ModelDeployment	A deployed model endpoint.
SglangRecipe	SGLang serving configuration.
VllmRecipe	vLLM serving configuration.