Training Configuration

Training Configuration

Settings below are taken from the DistilBERT source, not proposed defaults.

Setting	Full fine-tuning	Frozen backbone
Base model	distilbert-base-uncased	same
Epochs	4	6
Train batch size	8/device	8/device
Evaluation batch size	8/device	8/device
Learning rate	2e-5	1e-3
Warmup fraction	0.1	0.1
Weight decay	0.01	library default (not explicitly set)
Evaluation	Each epoch	Each epoch
Save	Each epoch, up to 2 retained	save_strategy='no'
Metric	f1_weighted for best checkpoint	evaluation comparison
Seed	42	42

The source chooses BF16 on a supported CUDA GPU; otherwise FP16 on CUDA; CPU uses neither. create_training_arguments and create_trainer introspect installed Transformers signatures to accommodate some version differences; compatibility across every future library version is not guaranteed.

Model downloads, installed package versions, and accelerator availability may affect real execution. No benchmark or GPU throughput result is claimed.