from transformers import AutoTokenizer, AutoModelForMaskedLM
tokenizer = AutoTokenizer.from_pretrained('ku-nlp/deberta-v2-tiny-japanese')
model = AutoModelForMaskedLM.from_pretrained('ku-nlp/deberta-v2-tiny-japanese')

sentence = '京都 大学 で 自然 言語 処理 を [MASK] する 。'  # input should be segmented into words by Juman++ in advance
encoding = tokenizer(sentence, return_tensors='pt')
...

You can also fine-tune this model on downstream tasks.

Tokenization

The input text should be segmented into words by Juman++ in advance. Juman++ 2.0.0-rc3 was used for pre-training. Each word is tokenized into subwords by sentencepiece.

Training data

We used the following corpora for pre-training:

Japanese Wikipedia (as of 20221020, 3.2GB, 27M sentences, 1.3M documents)
Japanese portion of CC-100 (85GB, 619M sentences, 66M documents)
Japanese portion of OSCAR (54GB, 326M sentences, 25M documents)

Note that we filtered out documents annotated with "header", "footer", or "noisy" tags in OSCAR. Also note that Japanese Wikipedia was duplicated 10 times to make the total size of the corpus comparable to that of CC-100 and OSCAR. As a result, the total size of the training data is 171GB.

Training procedure

We first segmented texts in the corpora into words using Juman++. Then, we built a sentencepiece model with 32000 tokens including words (JumanDIC) and subwords induced by the unigram language model of sentencepiece.

We tokenized the segmented corpora into subwords using the sentencepiece model and trained the Japanese DeBERTa model using transformers library. The training took 33 hours using 8 NVIDIA A100-SXM4-40GB GPUs.

The following hyperparameters were used during pre-training:

learning_rate: 1e-3
per_device_train_batch_size: 128
distributed_type: multi-GPU
num_devices: 8
gradient_accumulation_steps: 6
total_train_batch_size: 6,144
max_seq_length: 512
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-06
lr_scheduler_type: linear schedule with warmup
training_steps: 100,000
warmup_steps: 10,000

The accuracy of the trained model on the masked language modeling task was 0.593. The evaluation set consists of 5,000 randomly sampled documents from each of the training corpora.

Acknowledgments

This work was supported by Joint Usage/Research Center for Interdisciplinary Large-scale Information Infrastructures (JHPCN) through General Collaboration Project no. jh221004, "Developing a Platform for Constructing and Sharing of Large-Scale Japanese Language Models". For training models, we used the mdx: a platform for the data-driven future.