---
language: ja
license: cc-by-sa-4.0
datasets:
- wikipedia
- cc100
- oscar
mask_token: "[MASK]"
widget:
- text: "[MASK] 大学 で 自然 言語 処理 を 学ぶ 。"
---

# nlp-waseda/roberta-base-japanese

## Model description

This is a Japanese BigBird base model pretrained 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:
```python
from transformers import AutoTokenizer, AutoModelForMaskedLM
tokenizer = AutoTokenizer.from_pretrained("nlp-waseda/bigbird-base-japanese")
model = AutoModelForMaskedLM.from_pretrained("nlp-waseda/bigbird-base-japanese")

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

You can fine-tune this model on downstream tasks.

## Tokenization

The input text should be segmented into words by [Juman++](https://github.com/ku-nlp/jumanpp) in advance. Juman++ 2.0.0-rc3 was used for pretraining. Each word is tokenized into tokens by [sentencepiece](https://github.com/google/sentencepiece).

## Vocabulary

The vocabulary consists of 32000 tokens including words ([JumanDIC](https://github.com/ku-nlp/JumanDIC)) and subwords induced by the unigram language model of [sentencepiece](https://github.com/google/sentencepiece).

## Training procedure

This model was trained on Japanese Wikipedia (as of 20221101), the Japanese portion of CC-100, and the and the Japanese portion of oscar. It took two weeks using 16 NVIDIA A100 GPUs using [transformers](https://github.com/huggingface/transformers) and [DeepSpeed](https://github.com/microsoft/DeepSpeed).

The following hyperparameters were used during pretraining:
- learning_rate: 1e-4
- per_device_train_batch_size: 6
- gradient_accumulation_steps: 2
- total_train_batch_size: 192
- max_seq_length: 4096
- training_steps: 600000
- warmup_steps: 10000
- bf16: True
- deepspeed: ds_config.json

## Performance on JGLUE

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](https://www.jstage.jst.go.jp/article/jnlp/30/1/30_63/_pdf/-char/ja).

For tasks other than MARC-ja, the maximum length is short, so the attention_type was set to "original_full" and fine-tuning was performed for tasks other than MARC-ja. For MARC-ja, both "block_sparse" and "original_full" were used.

| Model                         | MARC-ja(original_full)/acc | JSTS/pearson | JSTS/spearman | JNLI/acc | JSQuAD/EM | JSQuAD/F1 | JComQA/acc |
|-------------------------------|--------------|---------------|----------|-----------|-----------|------------|------------|
| Waseda RoBERTa base           |  0.965       | 0.913        | 0.876         | 0.905    | 0.853     | 0.916     | 0.853      |
| Waseda RoBERTa large (seq512) | 0.969       | 0.925        | 0.890         | 0.928    | 0.910     | 0.955     | 0.900      |
| BigBird base (original_full)  |0.959       | 0.888        | 0.846         | 0.896    | 0.884     | 0.933     | 0.787      |
| BigBird base (block_sparse) |0.959       | -        | -         | -    | -    | -     | -      |