<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->
wav2vec2-xls-r-300m-ar-10
This model is a fine-tuned version of MeshalAlamr/wav2vec2-xls-r-300m-ar-9 on the common_voice dataset. It achieves the following results on the evaluation set:
- Loss: 87.0172
- Wer: 0.2017
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: 64
- eval_batch_size: 16
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 256
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 30
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 56.409 | 4.71 | 400 | 81.8407 | 0.2151 |
| 84.2726 | 9.41 | 800 | 82.6777 | 0.2237 |
| 80.3604 | 14.12 | 1200 | 85.3856 | 0.2226 |
| 70.7446 | 18.82 | 1600 | 87.9551 | 0.2180 |
| 61.3713 | 23.53 | 2000 | 88.0419 | 0.2096 |
| 54.5011 | 28.24 | 2400 | 87.0172 | 0.2017 |
Framework versions
- Transformers 4.17.0
- Pytorch 1.11.0
- Datasets 1.18.4
- Tokenizers 0.11.6
