Model Card for Japanese DeBERTa V3 base
Model description
This is a Japanese DeBERTa V3 base model pre-trained on LLM-jp corpus v1.0.
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-v3-base-japanese')
model = AutoModelForMaskedLM.from_pretrained('ku-nlp/deberta-v3-base-japanese')
sentences = [
"京都大学で自然言語処理を研究する。",
"I research NLP at Kyoto University.",
'int main() { printf("Hello, world!"); return 0; }',
]
encodings = tokenizer(sentences, return_tensors='pt')
...
You can also fine-tune this model on downstream tasks.
Tokenization
The tokenizer of this model is based on huggingface/tokenizers Unigram byte-fallback model.
The vocabulary entries were converted from llm-jp-tokenizer v2.2 (100k)
.
Please refer to README.md of llm-jp/llm-ja-tokenizer
for details on the vocabulary construction procedure.
Note that, unlike ku-nlp/deberta-v2-base-japanese, pre-segmentation by a morphological analyzer (e.g., Juman++) is no longer required for this model.
Training data
We used the LLM-jp corpus v1.0.1 for pre-training. The corpus consists of the following corpora:
- Japanese
- Wikipedia (1B tokens)
- mC4 (129B tokens)
- English
- Wikipedia (4B tokens)
- The Pile (126B tokens)
- Code
- The Stack (10B tokens)
We shuffled the corpora, which has 270B tokens in total, and trained the model for 2 epochs. Thus, the total number of tokens fed to the model was 540B.
Training procedure
We slightly modified the official implementation of DeBERTa V3 and followed the official training procedure. The modified code is available at nobu-g/DeBERTa.
The following hyperparameters were used during pre-training:
- learning_rate: 1e-4
- per_device_train_batch_size: 800
- num_devices: 8
- gradient_accumulation_steps: 3
- total_train_batch_size: 2400
- 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: 475,000
- warmup_steps: 10,000
Fine-tuning on NLU tasks
We fine-tuned the following models and evaluated them on the dev set of JGLUE. We tuned the learning rate and training epochs for each model and task following the JGLUE paper.
Model | MARC-ja/acc | JCoLA/acc | JSTS/pearson | JSTS/spearman | JNLI/acc | JSQuAD/EM | JSQuAD/F1 | JComQA/acc |
---|---|---|---|---|---|---|---|---|
Waseda RoBERTa base | 0.965 | 0.867 | 0.913 | 0.876 | 0.905 | 0.853 | 0.916 | 0.853 |
Waseda RoBERTa large (seq512) | 0.969 | 0.849 | 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 | 0.879 | 0.922 | 0.886 | 0.922 | 0.899 | 0.951 | 0.873 |
DeBERTaV2 large | 0.968 | 0.882 | 0.925 | 0.892 | 0.924 | 0.912 | 0.959 | 0.890 |
DeBERTaV3 base | 0.960 | 0.878 | 0.927 | 0.891 | 0.927 | 0.896 | 0.947 | 0.875 |
*The scores of LUKE are from the official repository.
License
Author
Nobuhiro Ueda (ueda at nlp.ist.i.kyoto-u.ac.jp)
Acknowledgments
This work was supported by Joint Usage/Research Center for Interdisciplinary Large-scale Information Infrastructures (JHPCN) through General Collaboration Project no. jh231006, "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.
- Downloads last month
- 1,159