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.