torgo_tiny_finetune_M02

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

• Loss: 0.3391
• Wer: 30.4754

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: 0.0001
• train_batch_size: 16
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 20

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.6178	0.85	500	0.3570	29.2869
0.105	1.7	1000	0.3471	35.7385
0.1006	2.55	1500	0.3797	35.1443
0.0661	3.4	2000	0.3132	49.8302
0.0483	4.25	2500	0.3368	62.6486
0.0335	5.1	3000	0.2921	39.7284
0.0271	5.95	3500	0.3178	31.8336
0.0222	6.8	4000	0.3214	56.6214
0.0188	7.65	4500	0.3255	29.3718
0.0135	8.5	5000	0.3525	40.3226
0.0098	9.35	5500	0.3004	31.3243
0.0094	10.2	6000	0.3255	29.5416
0.0063	11.05	6500	0.3111	32.3430
0.0042	11.9	7000	0.3198	42.1053
0.0027	12.76	7500	0.2946	26.9100
0.0028	13.61	8000	0.3201	32.0034
0.0015	14.46	8500	0.3236	31.0696
0.0008	15.31	9000	0.3244	29.9660
0.0004	16.16	9500	0.3332	31.8336
0.0004	17.01	10000	0.3586	30.3905
0.0001	17.86	10500	0.3415	29.6265
0.0	18.71	11000	0.3403	29.7963
0.0001	19.56	11500	0.3391	30.4754

Framework versions

• Transformers 4.32.0
• Pytorch 2.1.0+cu121
• Datasets 2.14.7
• Tokenizers 0.13.3