Struct Trainer 

pub struct Trainer { /* private fields */ }

High-level trainer that bundles a model and training config.

Construct via Trainer::builder() and the TrainerBuilder struct. Users provide token sequences (Vec<Vec<u32>>) from their own tokenizer.

Implementations

impl Trainer

pub fn builder() -> TrainerBuilder

Returns a new TrainerBuilder with sensible defaults.

pub fn train_on_token_sequences_with_callback( &mut self, sequences: &[Vec<u32>], on_step: impl FnMut(usize, f32), ) -> Result<TrainingReport>

Trains the model on the provided token sequences.

Each inner Vec<u32> is a tokenized text sample. Use your own tokenizer to produce these sequences before calling this method.

The on_step callback is invoked after each optimizer step with (step_index, loss_value).

pub fn train_on_token_sequences( &mut self, sequences: &[Vec<u32>], ) -> Result<TrainingReport>

Convenience wrapper that trains without a per-step callback.

pub fn train_on_chat_sequences_with_callback( &mut self, chat_pairs: &[(Vec<u32>, Vec<u32>)], on_step: impl FnMut(usize, f32), ) -> Result<TrainingReport>

Trains the model on chat-style (prompt, response) token-ID pairs.

Each element of chat_pairs is (prompt_token_ids, response_token_ids). The model sees the full context but loss is computed only on response tokens, preventing the model from learning to generate role labels.

The on_step callback is invoked after each optimizer step with (step_index, loss_value).

pub fn train_on_chat_sequences( &mut self, chat_pairs: &[(Vec<u32>, Vec<u32>)], ) -> Result<TrainingReport>

Convenience wrapper that trains on chat pairs without a per-step callback.

pub fn save_checkpoint(&self, path: &str) -> Result<()>

Saves a model checkpoint to the given path.

pub fn model(&self) -> &DefaultMultiscreenModel

Returns a reference to the underlying model.

pub fn model_mut(&mut self) -> &mut DefaultMultiscreenModel

Returns a mutable reference to the underlying model.

pub fn training_config(&self) -> &ModelTrainingConfig

Returns a reference to the training configuration.

Trait Implementations

impl Debug for Trainer

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.