torgo_tiny_finetune_M01

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: 0.3526
• Wer: 96.6044

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
0.6272	0.85	500	0.2872	24.1087
0.1055	1.7	1000	0.3364	77.5042
0.0998	2.55	1500	0.3646	65.8744
0.0635	3.4	2000	0.3276	34.9745
0.0521	4.24	2500	0.3619	31.8336
0.0368	5.09	3000	0.3158	43.0390
0.0269	5.94	3500	0.3424	53.7351
0.0215	6.79	4000	0.2886	48.8964
0.0182	7.64	4500	0.3331	31.0696
0.0135	8.49	5000	0.3308	45.0764
0.0092	9.34	5500	0.2825	28.9474
0.0088	10.19	6000	0.3169	32.3430
0.0056	11.04	6500	0.3223	55.7725
0.0034	11.88	7000	0.3396	30.2207
0.0041	12.73	7500	0.3403	31.8336
0.0031	13.58	8000	0.3544	138.4550
0.0023	14.43	8500	0.3357	54.8387
0.0004	15.28	9000	0.3618	53.6503
0.0003	16.13	9500	0.3598	74.3633
0.0002	16.98	10000	0.3536	98.8964
0.0003	17.83	10500	0.3529	95.8404
0.0001	18.68	11000	0.3505	98.0475
0.0001	19.52	11500	0.3526	96.6044

Framework versions

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