torgo_tiny_finetune_M05

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: 0.3072
• Wer: 24.1935

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
0.6262	0.84	500	0.3083	50.8489
0.1041	1.68	1000	0.3578	41.3413
0.0997	2.53	1500	0.3539	51.6978
0.0694	3.37	2000	0.3092	83.9559
0.0489	4.21	2500	0.3775	64.3463
0.0382	5.05	3000	0.3589	67.7419
0.0268	5.89	3500	0.3005	29.7114
0.0209	6.73	4000	0.3221	21.5620
0.0173	7.58	4500	0.3337	42.9542
0.0128	8.42	5000	0.3374	27.0798
0.011	9.26	5500	0.3639	20.7131
0.0083	10.1	6000	0.3622	24.9576
0.0066	10.94	6500	0.2958	21.5620
0.005	11.78	7000	0.3478	46.4346
0.0023	12.63	7500	0.3206	33.1919
0.0026	13.47	8000	0.3023	27.8438
0.0017	14.31	8500	0.2990	19.9491
0.0008	15.15	9000	0.2862	17.6570
0.0007	15.99	9500	0.2924	20.5433
0.0002	16.84	10000	0.2935	23.1749
0.0001	17.68	10500	0.3048	23.5993
0.0001	18.52	11000	0.3061	24.5331
0.0001	19.36	11500	0.3072	24.1935

Framework versions

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