nijatzeynalov's picture
Update README.md
5167bef
|
raw
history blame
2.2 kB
metadata
license: cc-by-nc-4.0
base_model: facebook/mms-1b-all
inference: false
tags:
  - generated_from_trainer
datasets:
  - common_voice_15_0
metrics:
  - wer
model-index:
  - name: wav2vec2-large-mms-1b-azerbaijani-common_voice15.0
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: common_voice_15_0
          type: common_voice_15_0
          config: az
          split: test
          args: az
        metrics:
          - name: Wer
            type: wer
            value: 0.2631578947368421

wav2vec2-large-mms-1b-azerbaijani-common_voice15.0

This model is a fine-tuned version of facebook/mms-1b-all on the common_voice_15_0 dataset. It achieves the following results on the evaluation set:

  • Loss: 0.3188
  • Wer: 0.2632

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: 0.001
  • train_batch_size: 16
  • eval_batch_size: 8
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 100
  • num_epochs: 15
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
7.6471 2.0 10 7.6790 1.0658
5.6745 4.0 20 4.2727 1.0088
3.5016 6.0 30 3.1003 1.0
2.6223 8.0 40 1.8137 1.0439
1.3939 10.0 50 0.6549 0.3947
0.3696 12.0 60 0.3665 0.2719
0.2475 14.0 70 0.3188 0.2632

Framework versions

  • Transformers 4.35.2
  • Pytorch 2.1.1+cu121
  • Datasets 2.15.0
  • Tokenizers 0.15.0