Edit model card

Mandarin_naive

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice dataset. It achieves the following results on the evaluation set:

  • Loss: 0.4584
  • Wer: 0.3999

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: 0.0003
  • train_batch_size: 16
  • eval_batch_size: 8
  • seed: 42
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 30
  • mixed_precision_training: Native AMP

Training results

Training Loss Epoch Step Validation Loss Wer
4.8963 3.67 400 1.0645 0.8783
0.5506 7.34 800 0.5032 0.5389
0.2111 11.01 1200 0.4765 0.4712
0.1336 14.68 1600 0.4815 0.4511
0.0974 18.35 2000 0.4956 0.4370
0.0748 22.02 2400 0.4881 0.4235
0.0584 25.69 2800 0.4732 0.4193
0.0458 29.36 3200 0.4584 0.3999

Framework versions

  • Transformers 4.16.2
  • Pytorch 1.10.0+cu111
  • Datasets 1.18.3
  • Tokenizers 0.11.0
Downloads last month
2

Dataset used to train GleamEyeBeast/Mandarin_naive