whisper-small-dhivehi-v3

This model is a fine-tuned version of Serialtechlab/whisper-small-dhivehi-v3 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.0469
• Wer: 0.4243
• Cer: 0.1538

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: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• 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: 500
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.0555	0.1778	500	0.0611	3.4904	1.5343
0.0476	0.3556	1000	0.0594	2.6611	1.0895
0.0438	0.5333	1500	0.0573	2.6697	1.2302
0.0417	0.7111	2000	0.0561	1.6096	0.7055
0.0387	0.8889	2500	0.0538	0.4946	0.1877
0.0239	1.0665	3000	0.0562	0.6677	0.2525
0.0238	1.2443	3500	0.0539	0.6063	0.3227
0.0306	1.4220	4000	0.0511	0.4554	0.2011
0.029	1.5998	4500	0.0498	0.5114	0.2605
0.0289	1.7776	5000	0.0474	0.4729	0.2151
0.0287	1.9554	5500	0.0469	0.4243	0.1538
0.0166	2.1330	6000	0.0498	0.7129	0.3167
0.015	2.3108	6500	0.0498	0.7703	0.3581
0.0154	2.4885	7000	0.0494	0.6739	0.2582
0.0154	2.6663	7500	0.0497	0.9540	0.4275
0.0144	2.8441	8000	0.0493	1.1120	0.4540

Framework versions

• Transformers 4.57.3
• Pytorch 2.9.1+cu128
• Datasets 4.4.2
• Tokenizers 0.22.1