This model is tranied from scratch on Cantonese text. It is a BERT model with a large architecture (24-layer, 1024-hidden, 16-heads, 326M parameters).

The first training stage is to pre-train the model on 128 length sequences with a batch size of 512 for 1 epoch. the second stage is to continued pre-train the model on 512 length sequences with a batch size of 512 for one more epoch.

How to use

You can use this model directly with a pipeline for masked language modeling:

python

from transformers import pipeline

mask_filler = pipeline(
    "fill-mask",
    model="hon9kon9ize/bert-large-cantonese"
)

mask_filler("雞蛋六隻，糖呢就兩茶匙，仲有[MASK]橙皮添。")

; [{'score': 0.08160534501075745,
;   'token': 943,
;   'token_str': '個',
;   'sequence': '雞 蛋 六 隻 ， 糖 呢 就 兩 茶 匙 ， 仲 有 個 橙 皮 添 。'},
;  {'score': 0.06182105466723442,
;   'token': 1576,
;   'token_str': '啲',
;   'sequence': '雞 蛋 六 隻 ， 糖 呢 就 兩 茶 匙 ， 仲 有 啲 橙 皮 添 。'},
;  {'score': 0.04600336775183678,
;   'token': 1646,
;   'token_str': '嘅',
;   'sequence': '雞 蛋 六 隻 ， 糖 呢 就 兩 茶 匙 ， 仲 有 嘅 橙 皮 添 。'},
;  {'score': 0.03743772581219673,
;   'token': 3581,
;   'token_str': '橙',
;   'sequence': '雞 蛋 六 隻 ， 糖 呢 就 兩 茶 匙 ， 仲 有 橙 橙 皮 添 。'},
;  {'score': 0.031560592353343964,
;   'token': 5148,
;   'token_str': '紅',
;   'sequence': '雞 蛋 六 隻 ， 糖 呢 就 兩 茶 匙 ， 仲 有 紅 橙 皮 添 。'}]

Training hyperparameters

The following hyperparameters were used during first training:

Batch size: 512
Learning rate: 1e-4
Learning rate scheduler: linear decay
1 Epoch
Warmup ratio: 0.1

Loss plot on WanDB

The following hyperparameters were used during second training:

Batch size: 512
Learning rate: 5e-5
Learning rate scheduler: linear decay
1 Epoch
Warmup ratio: 0.1

Loss plot on WanDB