wav2vec2-large-xls-r-300m-pl-colab
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_10_0 dataset. It achieves the following results on the evaluation set:
- Loss: 0.1589
- Wer: 0.1338
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: 6e-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
- lr_scheduler_warmup_steps: 765
- num_epochs: 10
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 4.9573 | 1.31 | 1000 | 3.2143 | 1.0 |
| 1.0216 | 2.61 | 2000 | 0.2310 | 0.2064 |
| 0.2925 | 3.92 | 3000 | 0.1888 | 0.1689 |
| 0.2252 | 5.23 | 4000 | 0.1731 | 0.1507 |
| 0.1983 | 6.54 | 5000 | 0.1722 | 0.1446 |
| 0.1757 | 7.84 | 6000 | 0.1637 | 0.1367 |
| 0.1656 | 9.15 | 7000 | 0.1589 | 0.1338 |
Framework versions
- Transformers 4.21.2
- Pytorch 1.11.0+cu113
- Datasets 2.4.0
- Tokenizers 0.12.1