learning_rate: 0.0001
train_batch_size: 4
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 16
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 2000
num_epochs: 251
mixed_precision_training: Native AMP

Training Loss	Epoch	Step	Validation Loss
6.5379	10.0452	20000	0.9408
3.4360	20.0904	40000	0.8064
2.9475	30.1357	60000	0.7386
2.6666	40.1809	80000	0.7093
2.4730	50.2261	100000	0.6868
2.3229	60.2713	120000	0.6716
2.2039	70.3165	140000	0.6586
2.1075	80.3618	160000	0.6550
2.0190	90.4070	180000	0.6472
1.9435	100.4522	200000	0.6215
1.8781	110.4974	220000	0.6290
1.8177	120.5426	240000	0.6302
1.7646	130.5879	260000	0.6168
1.7101	140.6331	280000	0.6162
1.6629	150.6783	300000	0.5982
1.6214	160.7235	320000	0.5996
1.5817	170.7687	340000	0.6098
1.5453	180.8140	360000	0.5998

wenda2025/OBert

OBert