Whisper Small - Mongolian LoRA
This model is a fine-tuned version of openai/whisper-small on own custom dataset. It achieves the following results on the evaluation set:
- Loss: 0.2533
- Wer: 0.1907
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-4
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 32
- 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_steps: 500
- training_steps: 4000
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | Validation Loss | Wer |
|---|---|---|---|---|
| 1.3506 | 1.6892 | 500 | 0.3970 | 0.3962 |
| 0.7855 | 3.3784 | 1000 | 0.2958 | 0.3157 |
| 0.5229 | 5.0676 | 1500 | 0.2583 | 0.2628 |
| 0.2567 | 6.7568 | 2000 | 0.2361 | 0.2334 |
| 0.0620 | 8.4459 | 2500 | 0.2421 | 0.2158 |
| 0.0136 | 10.1351 | 3000 | 0.2488 | 0.1944 |
| 0.0348 | 11.8243 | 3500 | 0.2580 | 0.2053 |
| 0.0058 | 13.5135 | 4000 | 0.2533 | 0.1907 |
Framework versions
- PEFT 0.18.1
- Transformers 5.0.0
- Pytorch 2.10.0+cu128
- Datasets 4.8.3
- Tokenizers 0.22.2