torgo_tiny_finetune_M04

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: 0.3396
• Wer: 32.3430

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
0.6227	0.84	500	0.3265	33.1070
0.1027	1.69	1000	0.3360	32.6825
0.1001	2.53	1500	0.4014	60.0170
0.0686	3.37	2000	0.3220	40.3226
0.048	4.22	2500	0.3146	31.3243
0.0366	5.06	3000	0.3477	57.8098
0.0262	5.9	3500	0.3054	21.8166
0.0237	6.75	4000	0.3007	43.4635
0.0153	7.59	4500	0.2969	24.8727
0.0149	8.43	5000	0.3628	52.8014
0.0112	9.27	5500	0.3670	29.7963
0.0096	10.12	6000	0.3354	24.5331
0.007	10.96	6500	0.3464	57.0458
0.0052	11.8	7000	0.3246	30.1358
0.0037	12.65	7500	0.3677	50.7640
0.0021	13.49	8000	0.3359	34.0407
0.002	14.33	8500	0.3406	41.7657
0.0011	15.18	9000	0.3296	36.3328
0.0004	16.02	9500	0.3359	33.5314
0.0	16.86	10000	0.3381	40.2377
0.0003	17.71	10500	0.3388	35.1443
0.0	18.55	11000	0.3410	33.4465
0.0001	19.39	11500	0.3396	32.3430

Framework versions

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