Bounded pretraining configuration

Maturity: verified

The resolved application mapping contains exactly these top-level fields:

Field	Contract
experiment_name	explicit selected experiment identity
seed	non-negative integer
artifacts.output_root	non-empty filesystem location or gs:// URI
dataset	registry path, dataset name, and train/val/test split
dataloader	positive batch size and non-negative worker count
normalization	independent I/Q-amplitude and physical-metadata policies
patching	positive selected patch size and unique supported-size list
model	implementation key and implementation-owned parameters
objective	implementation key and implementation-owned parameters
optimizer	implementation key and implementation-owned parameters
training	positive maximum steps and a CPU or NVIDIA CUDA GPU, selected by the exact device literal cpu or cuda
checkpoint	safe directory name, positive cadence, optional bounded-resume URI

Unknown outer fields fail Pydantic validation.

Component selections

Component	Implementation	Parameters
model	tiny_causal_iq	patch_size, d_model, num_layers, num_heads, max_token_sequence_length
objective	next_patch_prediction	none
optimizer	adamw	positive learning_rate, non-negative weight_decay

Implementation keys must use lowercase snake case. Each registration uses its own strict Pydantic model and rejects unknown parameters.

Supported normalization

• I/Q amplitude: per_sample_joint_max_abs or none;

• physical metadata: log10 or none.

The two policies override independently. Both epsilon values must be finite and positive.

Unsupported configuration

The current application schema has no epoch, validation, scheduler, best-checkpoint, mixed-dataset, distributed, callback, tracker, evaluator, or sweep configuration. Provider launch fields live in a separate YAML file.

See Also

• Configuration

• Configuration API

• Shipped experiment