whisper-large-v2-ft-cy-2602

This model is a fine-tuned version of openai/whisper-large-v2 on the DewiBrynJones/preprocessed-whisper-btb-cv-cvad-wlga-ca-2602 dataset. It achieves the following results on the evaluation set:

• Loss: 0.3849
• Wer: 0.2877

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: 16
• eval_batch_size: 16
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 32
• optimizer: Use adamw_torch_fused with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 500
• training_steps: 8000
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.5788	0.2196	500	0.5981	0.4791
0.4955	0.4392	1000	0.5043	0.4414
0.4459	0.6588	1500	0.4560	0.3769
0.4073	0.8783	2000	0.4282	0.3615
0.3267	1.0979	2500	0.4104	0.3240
0.3165	1.3175	3000	0.3949	0.3311
0.3078	1.5371	3500	0.3930	0.3327
0.292	1.7567	4000	0.3772	0.3002
0.2846	1.9763	4500	0.3684	0.3034
0.2068	2.1959	5000	0.3816	0.2874
0.2021	2.4155	5500	0.3753	0.2876
0.1973	2.6350	6000	0.3708	0.2910
0.1961	2.8546	6500	0.3691	0.2861
0.1481	3.0742	7000	0.3849	0.2877

Framework versions

• Transformers 4.57.6
• Pytorch 2.10.0+cu128
• Datasets 4.5.0
• Tokenizers 0.22.2