Added training info in README.md
Browse files
README.md
CHANGED
|
@@ -25,4 +25,41 @@ model-index:
|
|
| 25 |
- name: Wer
|
| 26 |
type: wer
|
| 27 |
value: 7.514
|
| 28 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
- name: Wer
|
| 26 |
type: wer
|
| 27 |
value: 7.514
|
| 28 |
+
---
|
| 29 |
+
|
| 30 |
+
# Whisper Medium nl - GeoffVdr
|
| 31 |
+
|
| 32 |
+
This model is a fine-tuned version of [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) on the Common Voice 11.0 dataset.
|
| 33 |
+
|
| 34 |
+
## Model description
|
| 35 |
+
More information needed
|
| 36 |
+
## Intended uses & limitations
|
| 37 |
+
More information needed
|
| 38 |
+
## Training and evaluation data
|
| 39 |
+
- Training: Mozilla CommonVoice 11 Dutch train+validation set
|
| 40 |
+
- Evaluation: Mozilla CommonVoice 11 Dutch test set
|
| 41 |
+
## Training procedure
|
| 42 |
+
|
| 43 |
+
## Training Hyperparameters
|
| 44 |
+
- learning_rate: 1e-5
|
| 45 |
+
- train_batch_size: 8
|
| 46 |
+
- eval_batch_size: 8
|
| 47 |
+
- gradient_accumulation_steps: 2
|
| 48 |
+
- lr_scheduler_warmup_steps: 500
|
| 49 |
+
- training_steps: 12000
|
| 50 |
+
|
| 51 |
+
## Training Results
|
| 52 |
+
|
| 53 |
+
| Training Loss | Epoch | Step | WER |
|
| 54 |
+
|:-------------:|:-----:|:----:|:----:|
|
| 55 |
+
| 0.1111 | 0.39 | 1000 | 9.89 |
|
| 56 |
+
| 0.0884 | 0.78 | 2000 | 9.26 |
|
| 57 |
+
| 0.0362 | 1.17 | 3000 | 8.64 |
|
| 58 |
+
| 0.0359 | 1.56 | 4000 | 8.60 |
|
| 59 |
+
| 0.0375 | 1.95 | 5000 | 8.24 |
|
| 60 |
+
:
|
| 61 |
+
:
|
| 62 |
+
| 0.0015 | 4.68 | 12000| 7.51 |
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
### Framework versions
|