bert-base-cantonese

This model is a continue pre-train version of bert-base-chinese on Cantonese Common Crawl dataset with 198m tokens.

Model description

This model has extended 500 more Chinese characters which very common in Cantonese, such as 冧, 噉, 麪, 笪, 冚, 乸 etc.

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: 0.0001
• train_batch_size: 24
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 192
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1.0

Training results

Framework versions

• Transformers 4.35.0.dev0
• Pytorch 2.1.1+cu121
• Datasets 2.14.6
• Tokenizers 0.14.1