davidilag/whisper-large-no-is-fo-fo_parl-3k-steps

whisper-large-no-is-fo-fo_parl-60k-steps

This model is a fine-tuned version of davidilag/whisper-large-no-is-fo-100h-30k-steps on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.5757
• Wer: 26.1232
• Cer: 13.6139

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: 48
• eval_batch_size: 20
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 192
• 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: 300
• training_steps: 3000

Training results

Training Loss	Epoch	Step	Validation Loss	Wer	Cer
0.486	14.4928	500	0.5783	26.3128	13.5818
0.4159	28.9855	1000	0.5665	25.9526	13.4001
0.4045	43.4783	1500	0.5611	26.1232	13.4678
0.39	57.9710	2000	0.5983	26.0474	13.5141
0.3986	72.4638	2500	0.5706	26.1611	13.5533
0.3941	86.9565	3000	0.5757	26.1232	13.6139

Framework versions

• Transformers 4.46.2
• Pytorch 2.9.0+cu128
• Datasets 3.0.1
• Tokenizers 0.20.3