deberta-v2-base-japanese

Model Card for Japanese DeBERTa V2 base

Model description

This is a Japanese DeBERTa V2 base model pre-trained on Japanese Wikipedia, the Japanese portion of CC-100, and the Japanese portion of OSCAR.

How to use

You can use this model for masked language modeling as follows:

from transformers import AutoTokenizer, AutoModelForMaskedLM
tokenizer = AutoTokenizer.from_pretrained('ku-nlp/deberta-v2-base-japanese-with-auto-jumanpp', trust_remote_code=True)
model = AutoModelForMaskedLM.from_pretrained('ku-nlp/deberta-v2-base-japanese-with-auto-jumanpp')

sentence = '京都大学で自然言語処理を[MASK]する。'
encoding = tokenizer(sentence, return_tensors='pt')
...

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

Tokenization

The input text is internally segmented by Juman++ within DebertaV2JumanppTokenizer or DebertaV2JumanppTokenizerFast, so there's no need to segment it in advance. To use DebertaV2JumanppTokenizer or DebertaV2JumanppTokenizerFast, you need to install Juman++ 2.0.0-rc3 and rhoknp.

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 three weeks using 8 NVIDIA A100-SXM4-40GB GPUs.

The following hyperparameters were used during pre-training:

• learning_rate: 2e-4
• per_device_train_batch_size: 44
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 6
• total_train_batch_size: 2,112
• 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: 500,000
• warmup_steps: 10,000

The accuracy of the trained model on the masked language modeling task was 0.779. The evaluation set consists of 5,000 randomly sampled documents from each of the training corpora.

Fine-tuning on NLU tasks

We fine-tuned the following models and evaluated them on the dev set of JGLUE. We tuned learning rate and training epochs for each model and task following the JGLUE paper.

*The scores of LUKE are from the official repository.

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.