wav2vec2-large-xls-r-300m-greek

This model was trained from scratch on the common_voice dataset. It achieves the following results on the evaluation set:

• Loss: 0.4823
• Wer: 0.3338

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: 3e-05
• 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
• num_epochs: 30
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.0106	1.72	200	0.5519	0.3537
0.0249	3.45	400	0.5174	0.3465
0.0206	5.17	600	0.4721	0.3323
0.0221	6.89	800	0.4652	0.3373
0.0204	8.62	1000	0.4883	0.3389
0.0192	10.34	1200	0.4785	0.3389
0.0186	12.07	1400	0.4789	0.3378
0.0172	13.79	1600	0.4915	0.3347
0.0184	15.52	1800	0.4759	0.3440
0.0168	17.24	2000	0.4891	0.3371
0.0155	18.96	2200	0.4928	0.3394
0.0146	20.69	2400	0.4834	0.3357
0.0146	22.41	2600	0.4814	0.3362
0.0151	24.14	2800	0.4791	0.3345
0.0136	25.86	3000	0.4825	0.3356
0.0136	27.58	3200	0.4850	0.3351
0.0127	29.31	3400	0.4823	0.3338

Framework versions

• Transformers 4.11.3
• Pytorch 1.10.0+cu111
• Datasets 1.18.3
• Tokenizers 0.10.3