torgo_tiny_finetune_F04

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: 0.3499
• Wer: 26.6553

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
0.6368	0.85	500	0.3637	28.9474
0.11	1.69	1000	0.3521	36.5025
0.0969	2.54	1500	0.2911	46.3497
0.0679	3.39	2000	0.2895	27.0798
0.053	4.24	2500	0.3115	26.9949
0.0361	5.08	3000	0.2972	28.8625
0.0278	5.93	3500	0.3036	26.9100
0.0233	6.78	4000	0.3311	59.0832
0.0148	7.63	4500	0.3000	27.6740
0.0149	8.47	5000	0.3317	37.6061
0.0105	9.32	5500	0.2975	29.4567
0.0087	10.17	6000	0.3593	27.1647
0.0075	11.02	6500	0.2840	28.0985
0.004	11.86	7000	0.3760	26.7402
0.0039	12.71	7500	0.3477	33.4465
0.0029	13.56	8000	0.3595	26.0611
0.0022	14.41	8500	0.3429	29.5416
0.0013	15.25	9000	0.2967	24.0238
0.0004	16.1	9500	0.3539	28.4380
0.0003	16.95	10000	0.3646	25.1273
0.0001	17.8	10500	0.3638	25.4669
0.0001	18.64	11000	0.3502	26.3158
0.0001	19.49	11500	0.3499	26.6553

Framework versions

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