CrucibleIR.Training.Config (CrucibleIR v0.2.1)

Configuration for model training.

Defines hyperparameters, optimizer settings, and training options for a training run.

Fields

:id - Config identifier (required)
:model_ref - Reference to model to train (required)
:dataset_ref - Training dataset (required)
:epochs - Number of training epochs
:batch_size - Batch size
:learning_rate - Initial learning rate
:optimizer - Optimizer type
:loss_function - Loss function
:metrics - Metrics to track
:validation_split - Validation data ratio
:device - Compute device
:seed - Random seed
:mixed_precision - Use mixed precision
:gradient_clipping - Max gradient norm
:early_stopping - Early stopping config
:checkpoint_every - Checkpoint frequency
:options - Additional options

Examples

iex> config = %CrucibleIR.Training.Config{
...>   id: :train_gpt2,
...>   model_ref: %CrucibleIR.ModelRef{id: :gpt2},
...>   dataset_ref: %CrucibleIR.DatasetRef{name: :wikitext},
...>   epochs: 10,
...>   batch_size: 32
...> }
iex> config.epochs
10

Summary

Types

device()

loss()

optimizer()

t()

Types

device()

@type device() :: :cpu | :cuda | :mps | :tpu | atom()

loss()

@type loss() :: :cross_entropy | :mse | :mae | :bce | atom()

optimizer()

@type optimizer() :: :adam | :sgd | :adamw | :rmsprop | atom()

t()

@type t() :: %CrucibleIR.Training.Config{
  batch_size: pos_integer(),
  checkpoint_every: pos_integer() | nil,
  dataset_ref: CrucibleIR.DatasetRef.t(),
  device: device(),
  early_stopping: map() | nil,
  epochs: pos_integer(),
  gradient_clipping: float() | nil,
  id: atom(),
  learning_rate: float(),
  loss_function: loss(),
  metrics: [atom()],
  mixed_precision: boolean(),
  model_ref: CrucibleIR.ModelRef.t(),
  optimizer: optimizer(),
  options: map() | nil,
  seed: integer() | nil,
  validation_split: float() | nil
}