whisper-small-serbian-v3

This model is a fine-tuned version of openai/whisper-small on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.4611
• Wer: 20.9326

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH 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: 1250
• num_epochs: 20
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.3423	0.9234	500	0.3223	27.8919
0.2588	1.8458	1000	0.2685	24.6306
0.1833	2.7682	1500	0.2665	23.3260
0.1154	3.6907	2000	0.2713	23.0167
0.0787	4.6131	2500	0.2881	22.8452
0.0456	5.5355	3000	0.3161	22.7374
0.025	6.4580	3500	0.3318	22.6472
0.0151	7.3804	4000	0.3657	22.9566
0.0084	8.3029	4500	0.3786	22.1788
0.0063	9.2253	5000	0.3875	21.9738
0.0037	10.1477	5500	0.3932	22.0038
0.0031	11.0702	6000	0.4101	21.7705
0.0032	11.9935	6500	0.4152	21.8005
0.003	12.9160	7000	0.4176	21.7935
0.0015	13.8384	7500	0.4213	21.5230
0.0008	14.7608	8000	0.4377	21.3339
0.0005	15.6833	8500	0.4389	21.2225
0.0002	16.6057	9000	0.4445	21.0935
0.0003	17.5282	9500	0.4536	21.0404
0.0001	18.4506	10000	0.4589	20.9220
0.0001	19.3730	10500	0.4611	20.9326

Framework versions

• Transformers 4.51.3
• Pytorch 2.6.0+cu124
• Datasets 3.6.0
• Tokenizers 0.21.4