TrainConfig

from modal_training_gym.common.train import TrainConfig

Compose dataset, model, and recipe into one training entrypoint.

Fields

Field	Type	Default	Description
dataset	DatasetConfig		The training dataset. train() materializes it into the framework’s /data volume before training if it isn’t already present.
model	ModelConfig		The model to train. Carries model identity (model_name) and weight-download logic; weights are downloaded into the shared HuggingFace cache volume on first use and reused across runs.
recipe	modal_training_gym.train_recipes.base.BaseTrainRecipe		Framework recipe (SlimeRecipe or MilesConfig). Selects the training framework and carries Modal infra settings (GPU type, node count, image) plus framework CLI flags.
checkpoint	modal_training_gym.common.checkpoint.Checkpoint | None	None	Checkpoint to resume training from. When None, training starts from the base model weights. Default None.
merge_model_recipe	bool	True	When True, merges the known-model preset recipe (e.g. Qwen3_4b_Recipe) onto recipe fields you left unset. Set False to run the recipe exactly as written, with no preset defaults. Default True.
detach	bool	True	Run the training app detached so it keeps running on Modal even if the local client disconnects (terminal closed, laptop asleep). Set False for an attached run that stops on Ctrl-C. Default True.
group_id	str | None	None	Shared sweep id. Set by TrainingGroup so every run in a sweep carries the same id, letting the dashboard group variants together. Not usually set by hand. Default None.
group_overrides	dict[str, Any] | None	None	Per-variant parameter overrides applied by TrainingGroup, keyed by dotted field path (e.g. {"recipe.lr": 1e-5}). Recorded in run metadata so the dashboard can label each variant. Default None.
group_axes	list[str] | None	None	Names of the swept parameter paths in a TrainingGroup grid. Recorded in run metadata for dashboard grouping; falls back to the keys of group_overrides when unset. Default None.

Methods

context_plan_line(self) -> str | None

One-line summary of the effective training context length and parallelism plan.

launch(self, *, show_output: bool = True, prepare_inputs: bool = False) -> modal_training_gym.common.run.TrainingRun

Start training in a detached Modal app and return immediately.

train(self, *, show_output: bool = True) -> modal_training_gym.common.train_result.TrainResult

Build the app, run training, and return the TrainResult.

Related Tutorials

• Qwen3-4B haiku evaluation with verifiable rewards — serve, evaluate, train, compare
• Code RL with Harbor hello-world and sandboxed verification
• Multi-turn number-guessing RL with custom generate and reward functions
• On-policy distillation on math — Qwen3-8B teacher, Qwen3-4B student
• DAPO on math with Qwen3-4B
• Audio GRPO on Qwen3-ASR-1.7B — transcribe LibriSpeech, reward −WER
• Sweep hyperparameters across runs with TrainingGroup
• GUI grounding with Qwen3-VL-8B — predict click coordinates from screenshots
• Cross-tokenizer agentic distillation on BFCL v3 multi-turn with live, execution-grounded rewards — DeepSeek V4 Flash teacher, Qwen3.6-35B-A3B student

Source: modal_training_gym/common/train.py