Struct Batch 

pub struct Batch {
    pub input_ids: Vec<Vec<u32>>,
    pub target_ids: Vec<Vec<u32>>,
    pub padded_input_ids: Vec<Vec<u32>>,
    pub padded_target_ids: Vec<Vec<u32>>,
    pub attention_mask: Vec<Vec<u8>>,
    pub loss_masks: Vec<Vec<u8>>,
    pub batch_size: usize,
    pub max_seq_len: usize,
}

Batch of training data

Fields

input_ids: Vec<Vec<u32>>

Unpadded inputs, one per sequence.

target_ids: Vec<Vec<u32>>

Unpadded targets, one per sequence.

padded_input_ids: Vec<Vec<u32>>

Inputs right-padded to max_seq_len with the pad token.

padded_target_ids: Vec<Vec<u32>>

Targets right-padded to max_seq_len with the pad token.

attention_mask: Vec<Vec<u8>>

1 for real tokens, 0 for padding.

loss_masks: Vec<Vec<u8>>

Per-sample loss masks aligned with padded targets (1 = train, 0 = ignore).

batch_size: usize

Number of sequences.

max_seq_len: usize

Length of the longest sequence.