learning_rate: 5e-06
train_batch_size: 4
eval_batch_size: 4
seed: 43
distributed_type: multi-GPU
gradient_accumulation_steps: 8
total_train_batch_size: 32
optimizer: Use OptimizerNames.ADAMW_BNB 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: 500
training_steps: 10000
mixed_precision_training: Native AMP

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.6068	0.6840	500	0.8540	0.5607	0.1909
0.2786	1.3680	1000	0.7612	0.5285	0.1891
0.1553	2.0520	1500	0.7578	0.4856	0.1657
0.1134	2.7360	2000	0.7510	0.4728	0.1777
0.0795	3.4200	2500	0.7529	0.4607	0.1621
0.0603	4.1040	3000	0.7617	0.4648	0.1519
0.0475	4.7880	3500	0.7778	0.4562	0.1496
0.0403	5.4720	4000	0.7808	0.4531	0.1619
0.0406	6.1560	4500	0.7930	0.4547	0.1508
0.0352	6.8399	5000	0.8042	0.4587	0.1516
0.0265	7.5239	5500	0.8057	0.5366	0.1769
0.0269	8.2079	6000	0.8101	0.4520	0.1489
0.0236	8.8919	6500	0.8130	0.4826	0.1843
0.0243	9.5759	7000	0.8301	0.4542	0.1542
0.0197	10.2599	7500	0.8340	0.4587	0.1544
0.0194	10.9439	8000	0.8496	0.4384	0.1479
0.0151	11.6279	8500	0.8418	0.4475	0.1520
0.017	12.3119	9000	0.8467	0.4381	0.1496
0.0173	12.9959	9500	0.8662	0.4833	0.1751
0.0158	13.6799	10000	0.8867	0.4739	0.1818

Mihaj/whisper-small-karelian-cs

whisper-small-karelian-cs