torgo_tiny_finetune_F03

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: 0.0640
• Wer: 15.0892

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	0.1136	7.8189
0.11	1.69	1000	0.0872	9.1907
0.0969	2.54	1500	0.0843	9.3278
0.0679	3.39	2000	0.0980	7.1331
0.053	4.24	2500	0.0756	7.1331
0.0361	5.08	3000	0.0637	9.1907
0.0278	5.93	3500	0.0491	8.3676
0.0233	6.78	4000	0.0446	27.8464
0.0148	7.63	4500	0.0403	12.8944
0.0149	8.47	5000	0.0748	28.6694
0.0105	9.32	5500	0.0631	17.6955
0.0087	10.17	6000	0.0619	12.0713
0.0075	11.02	6500	0.0525	18.6557
0.004	11.86	7000	0.0588	19.7531
0.0039	12.71	7500	0.0618	24.5542
0.0029	13.56	8000	0.0915	13.7174
0.0022	14.41	8500	0.0638	20.4390
0.0013	15.25	9000	0.0946	14.5405
0.0004	16.1	9500	0.0746	15.7750
0.0003	16.95	10000	0.0633	11.2483
0.0001	17.8	10500	0.0645	12.7572
0.0001	18.64	11000	0.0631	14.4033
0.0001	19.49	11500	0.0640	15.0892

Framework versions

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