wav2vec2-large-xls-r-300m-kurdish
This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - KMR dataset. It achieves the following results on the evaluation set:
- Loss: 0.2548
- Wer: 0.2688
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: 7e-05
- train_batch_size: 32
- eval_batch_size: 1
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 2000
- num_epochs: 100.0
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 1.3161 | 12.27 | 2000 | 0.4199 | 0.4797 |
| 1.0643 | 24.54 | 4000 | 0.2982 | 0.3721 |
| 0.9718 | 36.81 | 6000 | 0.2762 | 0.3333 |
| 0.8772 | 49.08 | 8000 | 0.2586 | 0.3051 |
| 0.8236 | 61.35 | 10000 | 0.2575 | 0.2865 |
| 0.7745 | 73.62 | 12000 | 0.2603 | 0.2816 |
| 0.7297 | 85.89 | 14000 | 0.2539 | 0.2727 |
| 0.7079 | 98.16 | 16000 | 0.2554 | 0.2681 |
Framework versions
- Transformers 4.16.0.dev0
- Pytorch 1.10.1+cu102
- Datasets 1.17.1.dev0
- Tokenizers 0.11.0