<!-- 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. -->
filipino-wav2vec2-l-xls-r-300m-test
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the filipino_voice dataset. It achieves the following results on the evaluation set:
- Loss: 0.7753
- Wer: 0.4831
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: 0.001
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 16
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 20
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 2.7314 | 2.09 | 400 | 0.7541 | 0.7262 |
| 0.6065 | 4.19 | 800 | 0.6738 | 0.6314 |
| 0.4063 | 6.28 | 1200 | 0.6310 | 0.5992 |
| 0.2986 | 8.38 | 1600 | 0.6301 | 0.5340 |
| 0.2263 | 10.47 | 2000 | 0.6598 | 0.5391 |
| 0.1714 | 12.57 | 2400 | 0.7778 | 0.5593 |
| 0.1303 | 14.66 | 2800 | 0.7231 | 0.4907 |
| 0.1056 | 16.75 | 3200 | 0.8031 | 0.4885 |
| 0.0851 | 18.85 | 3600 | 0.7753 | 0.4831 |
Framework versions
- Transformers 4.11.3
- Pytorch 1.10.0+cu111
- Datasets 1.18.3
- Tokenizers 0.10.3
