whisper-small-luganda-radio-mh
This model is a fine-tuned version of sulaimank/whisper-small-swahili-400hr on the None dataset. It achieves the following results on the evaluation set:
- Loss: 1.1962
- Wer: 0.5717
- Cer: 0.2111
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: 5e-05
- train_batch_size: 8
- eval_batch_size: 4
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 16
- optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: linear
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 20
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
|---|---|---|---|---|---|
| 3.9846 | 1.0 | 1234 | 1.4073 | 0.7990 | 0.2925 |
| 1.1773 | 2.0 | 2468 | 0.9149 | 0.6521 | 0.2332 |
| 0.7181 | 3.0 | 3702 | 0.7786 | 0.5410 | 0.1875 |
| 0.4213 | 4.0 | 4936 | 0.8041 | 0.5606 | 0.1972 |
| 0.2192 | 5.0 | 6170 | 0.9955 | 0.5560 | 0.1933 |
| 0.1088 | 6.0 | 7404 | 1.1962 | 0.5717 | 0.2111 |
Framework versions
- Transformers 4.57.1
- Pytorch 2.9.1+cu128
- Datasets 3.6.0
- Tokenizers 0.22.1