patrickvonplaten
/

wav2vec2-large-xls-r-300m-common_voice-tr-ft

@@ -20,9 +20,9 @@ should probably proofread and complete it, then remove this comment. -->
 This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the COMMON_VOICE - TR dataset.
 It achieves the following results on the evaluation set:
-- Loss: 0.3644
-- Wer: 0.3394
-- Cer: 0.0811
 ## Model description
@@ -41,45 +41,29 @@ More information needed
 ### Training hyperparameters
 The following hyperparameters were used during training:
-- learning_rate: 0.0003
-- train_batch_size: 4
 - eval_batch_size: 8
 - seed: 42
 - distributed_type: multi-GPU
 - num_devices: 8
-- total_train_batch_size: 32
 - total_eval_batch_size: 64
 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
 - lr_scheduler_type: linear
 - lr_scheduler_warmup_steps: 500
-- num_epochs: 100.0
 - mixed_precision_training: Native AMP
 ### Training results
-| Training Loss | Epoch | Step  | Validation Loss | Wer    | Cer    |
-|:-------------:|:-----:|:-----:|:---------------:|:------:|:------:|
-| 1.6613        | 4.59  | 500   | 0.8079          | 0.8542 | 0.2504 |
-| 1.3496        | 9.17  | 1000  | 0.4729          | 0.5968 | 0.1518 |
-| 1.1003        | 13.76 | 1500  | 0.4106          | 0.5225 | 0.1357 |
-| 1.1532        | 18.35 | 2000  | 0.3957          | 0.4978 | 0.1256 |
-| 1.0305        | 22.94 | 2500  | 0.3764          | 0.5008 | 0.1291 |
-| 0.8303        | 27.52 | 3000  | 0.3826          | 0.5113 | 0.1292 |
-| 0.9115        | 32.11 | 3500  | 0.3819          | 0.4324 | 0.1070 |
-| 0.8193        | 36.7  | 4000  | 0.3694          | 0.4223 | 0.1036 |
-| 0.8948        | 41.28 | 4500  | 0.3714          | 0.4100 | 0.1005 |
-| 0.774         | 45.87 | 5000  | 0.3558          | 0.3923 | 0.0971 |
-| 0.8194        | 50.46 | 5500  | 0.3729          | 0.4603 | 0.1180 |
-| 0.8616        | 55.05 | 6000  | 0.3616          | 0.3908 | 0.0963 |
-| 0.7901        | 59.63 | 6500  | 0.3575          | 0.3837 | 0.0952 |
-| 0.778         | 64.22 | 7000  | 0.3732          | 0.3790 | 0.0928 |
-| 0.7238        | 68.81 | 7500  | 0.3674          | 0.3734 | 0.0904 |
-| 0.6985        | 73.39 | 8000  | 0.3627          | 0.3615 | 0.0863 |
-| 0.5889        | 77.98 | 8500  | 0.3705          | 0.3548 | 0.0858 |
-| 0.5447        | 82.57 | 9000  | 0.3678          | 0.3534 | 0.0854 |
-| 0.4763        | 87.16 | 9500  | 0.3627          | 0.3509 | 0.0840 |
-| 0.3544        | 91.74 | 10000 | 0.3690          | 0.3495 | 0.0834 |
-| 0.4879        | 96.33 | 10500 | 0.3683          | 0.3418 | 0.0820 |
 ### Framework versions

 This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the COMMON_VOICE - TR dataset.
 It achieves the following results on the evaluation set:
+- Loss: 0.3527
+- Wer: 0.3212
+- Cer: 0.0750
 ## Model description
 ### Training hyperparameters
 The following hyperparameters were used during training:
+- learning_rate: 0.0005
+- train_batch_size: 8
 - eval_batch_size: 8
 - seed: 42
 - distributed_type: multi-GPU
 - num_devices: 8
+- total_train_batch_size: 64
 - total_eval_batch_size: 64
 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
 - lr_scheduler_type: linear
 - lr_scheduler_warmup_steps: 500
+- num_epochs: 50.0
 - mixed_precision_training: Native AMP
 ### Training results
+| Training Loss | Epoch | Step | Validation Loss | Wer    | Cer    |
+|:-------------:|:-----:|:----:|:---------------:|:------:|:------:|
+| 0.7897        | 9.09  | 500  | 0.4873          | 0.6006 | 0.1491 |
+| 0.5166        | 18.18 | 1000 | 0.4058          | 0.4238 | 0.1070 |
+| 0.4012        | 27.27 | 1500 | 0.3716          | 0.3781 | 0.0935 |
+| 0.3851        | 36.36 | 2000 | 0.3565          | 0.3458 | 0.0836 |
+| 0.2212        | 45.45 | 2500 | 0.3580          | 0.3272 | 0.0766 |
 ### Framework versions