whisper_sanlish_mnx_lr5e6
This model is a fine-tuned version of openai/whisper-small on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.8534
- Wer: 43.8951
- Cer: 10.2362
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-06
- 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_steps: 200
- num_epochs: 25
- mixed_precision_training: Native AMP
Training results
| Training Loss | Epoch | Step | CER | Validation Loss | WER |
|---|---|---|---|---|---|
| 16.1536 | 1.0 | 94 | 40.3500 | 4.5547 | 104.4400 |
| 8.0157 | 2.0 | 188 | 24.6719 | 1.7042 | 86.4783 |
| 5.1169 | 3.0 | 282 | 19.1951 | 1.3434 | 72.5530 |
| 4.0548 | 4.0 | 376 | 16.5179 | 1.2305 | 64.9849 |
| 3.5518 | 5.0 | 470 | 15.2406 | 1.1282 | 60.4440 |
| 3.0430 | 6.0 | 564 | 14.3482 | 1.0582 | 56.1049 |
| 2.7010 | 7.0 | 658 | 13.3158 | 1.0206 | 52.2704 |
| 2.5312 | 8.0 | 752 | 12.8784 | 1.0134 | 51.1604 |
| 2.2781 | 9.0 | 846 | 12.3360 | 0.9902 | 49.9495 |
| 2.1107 | 10.0 | 940 | 12.0560 | 0.9559 | 49.0414 |
| 2.0311 | 11.0 | 1034 | 11.8285 | 0.9351 | 48.8396 |
| 1.9949 | 12.0 | 1128 | 11.6360 | 0.9369 | 47.2250 |
| 1.8562 | 13.0 | 1222 | 11.6185 | 0.9258 | 48.0323 |
| 1.7036 | 14.0 | 1316 | 11.0586 | 0.9070 | 46.0141 |
| 1.7850 | 15.0 | 1410 | 11.0061 | 0.9228 | 45.3078 |
| 1.7146 | 16.0 | 1504 | 10.8661 | 0.8976 | 45.1060 |
| 1.6118 | 17.0 | 1598 | 10.6737 | 0.8963 | 44.6014 |
| 1.5558 | 18.0 | 1692 | 10.5162 | 0.8756 | 44.5005 |
| 1.5527 | 19.0 | 1786 | 10.4812 | 0.8594 | 44.0969 |
| 1.5170 | 20.0 | 1880 | 10.3587 | 0.8570 | 44.0969 |
| 1.4383 | 21.0 | 1974 | 10.1662 | 0.8623 | 43.4914 |
| 1.4371 | 22.0 | 2068 | 10.1312 | 0.8496 | 43.1887 |
| 1.4151 | 23.0 | 2162 | 10.2187 | 0.8519 | 43.8951 |
| 1.3676 | 24.0 | 2256 | 10.2887 | 0.8538 | 43.9960 |
| 1.4520 | 25.0 | 2350 | 10.2362 | 0.8534 | 43.8951 |
Framework versions
- Transformers 5.15.0
- Pytorch 2.11.0+cu128
- Datasets 4.0.0
- Tokenizers 0.22.2
Training configuration
| Setting | Value |
|---|---|
| Base model | openai/whisper-small |
| Target transcription scheme | Sanlish (Roman-script Santali) |
| Learning rate | 5e-06 |
| Planned epochs | 25 |
| Warmup steps | 200 |
| Batch size (train / eval) | 8 / 4 (grad accumulation 2) |
| Precision | fp16 |
| Save/eval strategy | per epoch |
| Hub checkpoint strategy | checkpoint (full resumable state pushed every epoch) |
| Early stopping | patience=5 on eval WER |
| Train / val / test split | 1489 / 192 / 194 |
Final test results
| Metric | Score |
|---|---|
| Test WER | 41.16% |
| Test CER | 9.52% |