bert-base-cantonese

This model is a continuation of indiejoseph/bert-base-cantonese, a BERT-based model pre-trained on a substantial corpus of Cantonese text. The dataset was sourced from a variety of platforms, including news articles, social media posts, and web pages. The text was segmented into sentences containing 11 to 460 tokens per line. To ensure data quality, Minhash LSH was employed to eliminate near-duplicate sentences, resulting in a final dataset comprising 161,338,273 tokens. Training was conducted using the run_mlm.py script from the transformers library.

This continuous pre-training aims to expand the model's knowledge with more up-to-date Hong Kong and Cantonese text data. So we slightly overfit the model with higher learng rate and more epochs.

WandB

Usage

from transformers import pipeline

pipe = pipeline("fill-mask", model="hon9kon9ize/bert-base-cantonese")

pipe("香港特首係李[MASK]超")

# [{'score': 0.3057154417037964,
#   'token': 2157,
#   'token_str': '家',
#   'sequence': '香 港 特 首 係 李 家 超'},
#  {'score': 0.08251259475946426,
#   'token': 6631,
#   'token_str': '超',
#   'sequence': '香 港 特 首 係 李 超 超'},
# ...

pipe("我睇到由治及興帶嚟[MASK]好處")

# [{'score': 0.9563464522361755,
#   'token': 1646,
#   'token_str': '嘅',
#   'sequence': '我 睇 到 由 治 及 興 帶 嚟 嘅 好 處'},
#  {'score': 0.00982475932687521,
#   'token': 4638,
#   'token_str': '的',
#   'sequence': '我 睇 到 由 治 及 興 帶 嚟 的 好 處'},
# ...

Intended uses & limitations

This model is intended to be used for further fine-tuning on Cantonese downstream tasks.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 180
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 1440
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10.0

Framework versions

• Transformers 4.45.0
• Pytorch 2.4.1+cu121
• Datasets 2.20.0
• Tokenizers 0.20.0