wav2vecvanilla_load_best

This model is a fine-tuned version of facebook/wav2vec2-base-960h on an unknown dataset. It achieves the following results on the evaluation set:

• best model at 1400 steps, training loss 0.871600, validation loss 0.835038, evaluation WER 0.324867

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: 4
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 20

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.47	0.43	100	1.0582	0.4027
1.2584	0.85	200	0.9775	0.3719
1.1377	1.28	300	0.9759	0.3628
1.0955	1.71	400	1.1271	0.3626
1.0945	2.14	500	0.9063	0.3589
1.0567	2.56	600	0.9288	0.3564
1.033	2.99	700	1.1634	0.3522
1.0043	3.42	800	0.8396	0.3506
0.9776	3.85	900	0.8654	0.3437
0.9147	4.27	1000	0.8816	0.3362
0.9041	4.7	1100	0.8994	0.3303
0.8648	5.13	1200	0.8379	0.3361
0.8241	5.56	1300	0.8263	0.3292
0.8716	5.98	1400	0.8350	0.3249
0.8218	6.41	1500	nan	1.0
0.0	6.84	1600	nan	1.0
0.0	7.26	1700	nan	1.0
0.0	7.69	1800	nan	1.0
0.0	8.12	1900	nan	1.0
0.0	8.55	2000	nan	1.0
0.0	8.97	2100	nan	1.0
0.0	9.4	2200	nan	1.0
0.0	9.83	2300	nan	1.0
0.0	10.26	2400	nan	1.0
0.0	10.68	2500	nan	1.0
0.0	11.11	2600	nan	1.0
0.0	11.54	2700	nan	1.0
0.0	11.97	2800	nan	1.0
0.0	12.39	2900	nan	1.0
0.0	12.82	3000	nan	1.0
0.0	13.25	3100	nan	1.0
0.0	13.68	3200	nan	1.0
0.0	14.1	3300	nan	1.0
0.0	14.53	3400	nan	1.0
0.0	14.96	3500	nan	1.0
0.0	15.38	3600	nan	1.0
0.0	15.81	3700	nan	1.0
0.0	16.24	3800	nan	1.0
0.0	16.67	3900	nan	1.0
0.0	17.09	4000	nan	1.0
0.0	17.52	4100	nan	1.0
0.0	17.95	4200	nan	1.0
0.0	18.38	4300	nan	1.0
0.0	18.8	4400	nan	1.0
0.0	19.23	4500	nan	1.0
0.0	19.66	4600	nan	1.0

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2