lisbet.training.core#

Training and fitting functions for LISBET.

Notes

[a] The dictionary of RNG seed could be refactored as a Pydantic model in the future.

[b] The train/dev split is performed here and not in the input_pipeline module to

emphasize that the test set is frozen and won’t be used for hyper-parameters tuning.

[c] When mixing datasets of different lengths, the training and evaluation loops will

stop after exhausting the shortest one. Please consider using random sampling.

Functions

train(experiment_config)

Train a LISBET model.

lisbet.training.core.train(experiment_config)[source]#

Train a LISBET model.

This function orchestrates the full training pipeline for LISBET, including data loading, model construction, training, evaluation, and saving artifacts. All parameters match the CLI arguments exactly.

Parameters:

experiment_config (ExperimentConfig) – Configuration object containing all parameters for the training run. It includes data paths, model architecture, training hyperparameters, and task definitions. Must be a Pydantic model.

Returns:

model – The trained LISBET model instance.

Return type:

Module

Notes

All arguments are exposed for CLI and documentation. For advanced usage, see the LISBET documentation.