"Complete reference for the training-gym Python library.\n\n## Core\n\n| Class | Description |\n|-------|-------------|\n| [`ModelConfig`](/reference/core/modelconfig/) | Base class for model identity and weight-download logic. |\n| [`HFModelConfiguration`](/reference/core/hfmodelconfiguration/) | ModelConfig for models hosted on HuggingFace. |\n| [`ModelArchitecture`](/reference/core/modelarchitecture/) | Transformer architecture parameters for a specific model. |\n| [`DatasetConfig`](/reference/core/datasetconfig/) | Dataset configuration shared across training frameworks. |\n| [`HuggingFaceDataset`](/reference/core/huggingfacedataset/) | Dataset backed by a HuggingFace `datasets` repo. |\n| [`HarborDataset`](/reference/core/harbordataset/) | Dataset backed by a Harbor task directory structure. |\n| [`WandbConfig`](/reference/core/wandbconfig/) | Weights & Biases logging configuration shared across all frameworks. |\n| [`ModalRayCluster`](/reference/core/modalraycluster/) | Base class for bootstrapping a Ray cluster inside Modal clustered functions. |\n| [`TrainResult`](/reference/core/trainresult/) | One completed training run's checkpoint handle. |\n\n## Evaluation\n\n| Class | Description |\n|-------|-------------|\n| [`EvalConfig`](/reference/evaluation/evalconfig/) | Evaluate a deployed model on a dataset config. |\n| [`EvalResult`](/reference/evaluation/evalresult/) | Saved results for one evaluation run across a dataset. |\n| [`EvalRowResult`](/reference/evaluation/evalrowresult/) | One model interaction: the prompt, the raw response, its parsed |\n| [`HarborEval`](/reference/evaluation/harboreval/) | Evaluate a deployed model on a Harbor dataset using sandbox execution. |\n\n## Models\n\n| Class | Description |\n|-------|-------------|\n| [`ToolCall`](/reference/models/toolcall/) | A parsed tool invocation from model output. |\n| [`ParsedResponse`](/reference/models/parsedresponse/) | Structured result of parsing raw model output. |\n| [`parse_qwen3_response`](/reference/models/parse_qwen3_response/) | Parse Qwen3-family model output into structured content. |\n| [`Qwen3-0.6B`](/reference/models/qwen3_0_6b/) | Qwen3-0.6B (0.6 billion parameters) from Alibaba. |\n| [`Qwen3-1.7B`](/reference/models/qwen3_1_7b/) | Qwen3-1.7B (1.7 billion parameters) from Alibaba. |\n| [`Qwen3-4B`](/reference/models/qwen3_4b/) | Qwen3-4B (4 billion parameters) from Alibaba. |\n| [`Qwen3-8B`](/reference/models/qwen3_8b/) | Qwen3-8B (8 billion parameters) from Alibaba. |\n| [`Qwen3-30B-A3B`](/reference/models/qwen3_30b/) | Qwen3-30B-A3B (30B total, ~3B active) MoE model from Alibaba. |\n| [`Qwen3.6-35B-A3B`](/reference/models/qwen3_6_35b/) | Qwen3.6-35B-A3B (35B total, ~3B active) MoE model from Alibaba. |\n| [`Qwen3.6-27B`](/reference/models/qwen3_6_27b/) | Qwen3.6-27B dense hybrid Gated DeltaNet/attention model. |\n\n## Training\n\n| Class | Description |\n|-------|-------------|\n| [`TrainConfig`](/reference/training/trainconfig/) | Compose dataset, model, and recipe into one training entrypoint. |\n| [`TrainingGroup`](/reference/training/traininggroup/) | A base `TrainConfig` plus a grid of field overrides to expand. |\n| [`SlimeRecipe`](/reference/training/slimerecipe/) | Recipe dataclass for configuring slime GRPO training on Modal. |\n| [`Qwen3_6_35b_Recipe`](/reference/training/qwen3_6_35b_recipe/) | Qwen3.6-35B-A3B (MoE) on 1×8×H100 with TP2/PP2/CP1/EP4. |\n| [`Qwen3_6_27b_Recipe`](/reference/training/qwen3_6_27b_recipe/) | Qwen3.6-27B on 4×8×H100 using Slime's Qwen3.5-27B recipe. |\n\n## Deployment\n\n| Class | Description |\n|-------|-------------|\n| [`DeploymentConfig`](/reference/deployment/deploymentconfig/) | Deploy a model behind a serving engine. |\n| [`ModelDeployment`](/reference/deployment/modeldeployment/) | A deployed model endpoint. |\n| [`SglangRecipe`](/reference/deployment/sglangrecipe/) | SGLang serving configuration. |\n| [`VllmRecipe`](/reference/deployment/vllmrecipe/) | vLLM serving configuration. |\n"

API Reference

API reference for training-gym public classes.

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.
`Qwen3-30B-A3B`	Qwen3-30B-A3B (30B total, ~3B active) MoE model from Alibaba.
`Qwen3.6-35B-A3B`	Qwen3.6-35B-A3B (35B total, ~3B active) MoE model from Alibaba.
`Qwen3.6-27B`	Qwen3.6-27B dense hybrid Gated DeltaNet/attention model.

Training

Class	Description
`TrainConfig`	Compose dataset, model, and recipe into one training entrypoint.
`TrainingGroup`	A base `TrainConfig` plus a grid of field overrides to expand.
`SlimeRecipe`	Recipe dataclass for configuring slime GRPO training on Modal.
`Qwen3_6_35b_Recipe`	Qwen3.6-35B-A3B (MoE) on 1×8×H100 with TP2/PP2/CP1/EP4.
`Qwen3_6_27b_Recipe`	Qwen3.6-27B on 4×8×H100 using Slime’s Qwen3.5-27B recipe.

Deployment

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