wavLM-VLSP-vi

This model is a fine-tuned version of microsoft/wavlm-base-plus on the PHONGDTD/VINDATAVLSP - NA dataset. It achieves the following results on the evaluation set:

• Loss: 45.8892
• Wer: 0.9999
• Cer: 0.9973

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: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• total_train_batch_size: 8
• total_eval_batch_size: 16
• 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
3.4482	9.41	40000	3.4480	0.9999	0.9974
3.4619	18.81	80000	3.4514	0.9999	0.9974
3.7961	28.22	120000	3.8732	0.9999	0.9974
24.3843	37.62	160000	22.5457	0.9999	0.9973
48.5691	47.03	200000	45.8892	0.9999	0.9973

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.0+cu111
• Datasets 1.18.3
• Tokenizers 0.10.3