|
--- |
|
language: ja |
|
license: cc-by-sa-4.0 |
|
library_name: transformers |
|
tags: |
|
- deberta |
|
- deberta-v2 |
|
- fill-mask |
|
datasets: |
|
- wikipedia |
|
- cc100 |
|
- oscar |
|
metrics: |
|
- accuracy |
|
mask_token: "[MASK]" |
|
widget: |
|
- text: "京都 大学 で 自然 言語 処理 を [MASK] する 。" |
|
--- |
|
|
|
# Model Card for Japanese DeBERTa V2 large |
|
|
|
## Model description |
|
|
|
This is a Japanese DeBERTa V2 large 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: |
|
|
|
```python |
|
from transformers import AutoTokenizer, AutoModelForMaskedLM |
|
|
|
tokenizer = AutoTokenizer.from_pretrained('ku-nlp/deberta-v2-large-japanese') |
|
model = AutoModelForMaskedLM.from_pretrained('ku-nlp/deberta-v2-large-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++](https://github.com/ku-nlp/jumanpp) in |
|
advance. [Juman++ 2.0.0-rc3](https://github.com/ku-nlp/jumanpp/releases/tag/v2.0.0-rc3) was used for pre-training. Each |
|
word is tokenized into subwords by [sentencepiece](https://github.com/google/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++](https://github.com/ku-nlp/jumanpp). |
|
Then, we built a sentencepiece model with 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). |
|
|
|
We tokenized the segmented corpora into subwords using the sentencepiece model and trained the Japanese DeBERTa model |
|
using [transformers](https://github.com/huggingface/transformers) library. |
|
The training took 36 days using 8 NVIDIA A100-SXM4-40GB GPUs. |
|
|
|
The following hyperparameters were used during pre-training: |
|
|
|
- learning_rate: 1e-4 |
|
- per_device_train_batch_size: 18 |
|
- distributed_type: multi-GPU |
|
- num_devices: 8 |
|
- gradient_accumulation_steps: 16 |
|
- total_train_batch_size: 2,304 |
|
- 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: 300,000 |
|
- warmup_steps: 10,000 |
|
|
|
The accuracy of the trained model on the masked language modeling task was 0.799. |
|
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](https://www.jstage.jst.go.jp/article/jnlp/30/1/30_63/_pdf/-char/ja). |
|
|
|
| Model | MARC-ja/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 | |
|
| LUKE Japanese base* | 0.965 | 0.916 | 0.877 | 0.912 | - | - | 0.842 | |
|
| LUKE Japanese large* | 0.965 | 0.932 | 0.902 | 0.927 | - | - | 0.893 | |
|
| DeBERTaV2 base | 0.970 | 0.922 | 0.886 | 0.922 | 0.899 | 0.951 | 0.873 | |
|
| DeBERTaV2 large | 0.968 | 0.925 | 0.892 | 0.924 | 0.912 | 0.959 | 0.890 | |
|
|
|
*The scores of LUKE are from [the official repository](https://github.com/studio-ousia/luke). |
|
|
|
## 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. |
|
|