Struct TrainingExample 

pub struct TrainingExample {
    pub input_ids: Vec<u32>,
    pub target_ids: Vec<u32>,
    pub loss_mask: Vec<u8>,
}

One training example with optional response-only loss mask.

Fields

input_ids: Vec<u32>

Tokens fed to the model.

target_ids: Vec<u32>

Next-token targets, aligned with input_ids.

loss_mask: Vec<u8>

1 = compute loss, 0 = ignore. Same length as target_ids.

Implementations

impl TrainingExample

pub fn from_next_token(sequence: Vec<u32>) -> Self

Next-token example over a whole sequence: inputs are all but the last token, targets all but the first, every target supervised.

pub fn from_sft_prompt_response( prompt_tokens: Vec<u32>, response_tokens: Vec<u32>, ) -> Self

Build an SFT example where loss is only computed on response tokens.

prompt_tokens are the tokens up to (and including) the assistant prefix, response_tokens are the generated / supervised response ids.

Trait Implementations

impl Clone for TrainingExample

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for TrainingExample

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

Formats the value using the given formatter.