Training

The training key of the config.

Controls the model training process.

Field	Type	Required	Default	Description
max_epochs	integer (≥ 1)	yes		The maximum number of epochs to train for.
batch_size	integer (≥ 1)	yes		The number of inputs to process in each training batch (see: input_preprocessing).
optimizer	Optimizer	yes		Controls the optimization algorithm used to update the model's weights.
scheduler	Scheduler	yes		Controls how the learning rate is modified over the course of training.
encoder_training_strategy	EncoderTrainingStrategy	yes		Controls how encoder weights are managed during training.
early_stopping	EarlyStopping	no	null	End fine tuning early if model performance is no longer improving.

Optimizer

One of the following, selected by its name field:

name	Type
adamw	AdamW
manual	EscapeHatchComponent

Scheduler

One of the following, selected by its name field:

name	Type
manual	EscapeHatchComponent
plateau	Plateau

EncoderTrainingStrategy

One of the following, selected by its name field:

name	Type
freeze_unfreeze	EncoderFreezing
layer_wise_lr_decay	LayerWiseLRDecay

EarlyStopping

Controls early-stopping behavior during supervised fine tuning.

Monitors a validation metric and stops training after patience epochs without improvement. When monitoring begins depends on the encoder strategy: immediately for layer-wise LR decay, or after the encoder unfreeze epoch for freeze/unfreeze.

Field	Type	Required	Default	Description
patience	integer (≥ 1)	yes		Epochs without improvement before stopping.

AdamW

Field	Type	Required	Default	Description
name	"adamw"	yes
lr	number (≥ 0)	yes		The learning rate to use for the optimizer.

Plateau

Lowers learning rate via an exponential decay strategy when model performance plateaus.

Field	Type	Required	Default	Description
name	"plateau"	yes
factor	number (≥ 0, ≤ 1)	yes		Multiplicative factor of learning rate decay. LR = LR * factor.
patience	integer (≥ 0)	yes		Number of epochs with no improvement after which learning rate will be reduced.
min_lr	number (≥ 0)	yes		A lower bound on the learning rate.
cooldown	integer (≥ 0)	yes		Number of epochs to wait before resuming normal operation after lr has been reduced.

EncoderFreezing

Freeze encoder weights for a number of epochs, then unfreeze at a reduced learning rate.

Field	Type	Required	Default	Description
name	"freeze_unfreeze"	no	"freeze_unfreeze"
unfreeze_at_epoch	integer (≥ 0)	yes		The epoch at which weights in the encoder can be updated by the optimizer.
unfreeze_lr_factor	number (≥ 0)	yes		The factor by which the learning rate is multiplied when the encoder is unfrozen.

LayerWiseLRDecay

Apply per-layer learning rate decay to the encoder from the start of training.

Each encoder layer receives a learning rate scaled by decay_factor relative to the next layer. The configured optimizer learning rate applies to the decoder (last layer). Layer i gets: lr * decay_factor^(num_layers - i).

Field	Type	Required	Default	Description
name	"layer_wise_lr_decay"	yes
decay_factor	number (> 0, ≤ 1)	yes		Multiplicative factor applied per layer. Layer i LR = base_lr * decay_factor^(num_layers - i).
num_layers	integer (≥ 1)	yes		Number of encoder layers. Must match the foundation model metadata (e.g. 4 for Nano, 12 for Tiny/Base).