<!-- 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. -->
fine-tune-Wav2Vec2-XLS-R-300M-Indonesia-test
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: 0.3076
- Wer: 0.2971
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: 0.0001
- train_batch_size: 7
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 14
- 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 |
---|---|---|---|---|
2.9436 | 9.99 | 570 | 2.7467 | 1.0 |
1.0498 | 19.99 | 1140 | 0.3630 | 0.3965 |
0.6789 | 29.99 | 1710 | 0.3396 | 0.3712 |
0.5259 | 39.99 | 2280 | 0.3204 | 0.3241 |
0.4701 | 49.99 | 2850 | 0.3118 | 0.3005 |
0.4248 | 59.99 | 3420 | 0.3076 | 0.2971 |
Framework versions
- Transformers 4.23.1
- Pytorch 1.10.0+cu113
- Datasets 2.6.1
- Tokenizers 0.13.1