File size: 2,340 Bytes
2e17744 f2626e8 2e17744 f2626e8 2e17744 f2626e8 7245ff3 f2626e8 3fc01ce f2626e8 28df4e5 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 |
---
language: ja
license: cc-by-sa-4.0
datasets:
- wikipedia
- cc100
mask_token: "[MASK]"
widget:
- text: "早稲田 大学 で 自然 言語 処理 を [MASK] する 。"
---
# nlp-waseda/roberta-large-japanese
## Model description
This is a Japanese RoBERTa large model pretrained on Japanese Wikipedia and the Japanese portion of CC-100.
## 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/roberta-large-japanese")
model = AutoModelForMaskedLM.from_pretrained("nlp-waseda/roberta-large-japanese")
sentence = '早稲田 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 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).
`BertJapaneseTokenizer` now supports automatic `JumanppTokenizer` and `SentencepieceTokenizer`. You can use [this model](https://huggingface.co/nlp-waseda/roberta-large-japanese-with-auto-jumanpp) without any data preprocessing.
## 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 20210920) and the Japanese portion of CC-100. It took two weeks using eight NVIDIA A100 GPUs.
The following hyperparameters were used during pretraining:
- learning_rate: 6e-5
- per_device_train_batch_size: 103
- distributed_type: multi-GPU
- num_devices: 8
- gradient_accumulation_steps: 5
- total_train_batch_size: 4120
- max_seq_length: 128
- optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-6
- lr_scheduler_type: linear
- training_steps: 670000
- warmup_steps: 10000
- mixed_precision_training: Native AMP
## Performance on JGLUE
See the [Baseline Scores](https://github.com/yahoojapan/JGLUE#baseline-scores) of JGLUE.
|