torgo_tiny_finetune_M03

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: 0.3315
• Wer: 28.3531

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
0.6365	0.85	500	0.3228	25.3820
0.1068	1.71	1000	0.3417	52.2920
0.1009	2.56	1500	0.3318	48.2173
0.0686	3.41	2000	0.2947	30.3056
0.0516	4.27	2500	0.3396	26.0611
0.0353	5.12	3000	0.3153	28.4380
0.0255	5.97	3500	0.2689	24.7878
0.0207	6.83	4000	0.4144	35.1443
0.0148	7.68	4500	0.2552	31.8336
0.015	8.53	5000	0.3356	32.5127
0.0114	9.39	5500	0.3311	32.9372
0.0098	10.24	6000	0.3318	24.4482
0.0067	11.09	6500	0.2942	29.5416
0.0031	11.95	7000	0.3945	60.8659
0.0044	12.8	7500	0.3343	28.5229
0.0026	13.65	8000	0.3204	23.3447
0.0019	14.51	8500	0.3103	24.6180
0.001	15.36	9000	0.3257	29.6265
0.0004	16.21	9500	0.3615	28.1834
0.0001	17.06	10000	0.3410	26.7402
0.0002	17.92	10500	0.3327	28.1834
0.0001	18.77	11000	0.3314	28.6078
0.0001	19.62	11500	0.3315	28.3531

Framework versions

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