wav2vec2-large-xls-r-300m-gl-jupyter7

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.1004
• Wer: 0.0647

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: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 60
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.8074	3.36	400	0.4882	0.5245
0.2396	6.72	800	0.1335	0.1524
0.0876	10.08	1200	0.1216	0.1199
0.0597	13.44	1600	0.1289	0.1241
0.0449	16.8	2000	0.1164	0.1028
0.0372	20.17	2400	0.1270	0.1023
0.0319	23.53	2800	0.1111	0.0966
0.0286	26.89	3200	0.1142	0.0925
0.0246	30.25	3600	0.1142	0.0926
0.0235	33.61	4000	0.1075	0.0836
0.0181	36.97	4400	0.1083	0.0837
0.0151	40.33	4800	0.1140	0.0768
0.014	43.69	5200	0.1015	0.0748
0.0111	47.06	5600	0.1023	0.0702
0.0093	50.42	6000	0.1028	0.0708
0.0078	53.78	6400	0.0999	0.0645
0.0071	57.14	6800	0.1004	0.0647

Framework versions

• Transformers 4.11.3
• Pytorch 1.9.1+cu111
• Datasets 1.18.3
• Tokenizers 0.10.3