File size: 448 Bytes
f5d6a7c
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
"""Reduce training padding while retaining sequential validation batches."""

from torch.utils.data import SequentialSampler
from trl import SFTTrainer


class OptimizedSFTTrainer(SFTTrainer):
    def _get_eval_sampler(self, eval_dataset):
        # Training length grouping should not change validation's batch ordering.
        # Accelerate shards this sequential loader and gathers the full split.
        return SequentialSampler(eval_dataset)