Add model
Browse files- README.md +98 -1
- config.json +31 -0
- pytorch_model.bin +3 -0
- special_tokens_map.json +9 -0
- spm.model +3 -0
- tokenizer.json +0 -0
- tokenizer_config.json +15 -0
README.md
CHANGED
@@ -1,3 +1,100 @@
|
|
1 |
---
|
2 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
language: ja
|
3 |
+
license: cc-by-sa-4.0
|
4 |
+
library_name: transformers
|
5 |
+
datasets:
|
6 |
+
- cc100
|
7 |
+
- mc4
|
8 |
+
- oscar
|
9 |
+
- wikipedia
|
10 |
+
- izumi-lab/cc100-ja
|
11 |
+
- izumi-lab/mc4-ja-filter-ja-normal
|
12 |
+
- izumi-lab/oscar2301-ja-filter-ja-normal
|
13 |
+
- izumi-lab/wikipedia-ja-20230720
|
14 |
+
- izumi-lab/wikinews-ja-20230728
|
15 |
+
|
16 |
+
widget:
|
17 |
+
- text: 東京大学で[MASK]の研究をしています。
|
18 |
+
|
19 |
---
|
20 |
+
|
21 |
+
# DeBERTa V2 base Japanese
|
22 |
+
|
23 |
+
This is a [DeBERTaV2](https://github.com/microsoft/DeBERTa) model pretrained on Japanese texts.
|
24 |
+
The codes for the pretraining are available at [retarfi/language-pretraining](https://github.com/retarfi/language-pretraining/releases/tag/v2.2.1).
|
25 |
+
|
26 |
+
|
27 |
+
## How to use
|
28 |
+
|
29 |
+
You can use this model for masked language modeling as follows:
|
30 |
+
|
31 |
+
```python
|
32 |
+
from transformers import AutoTokenizer, AutoModelForMaskedLM
|
33 |
+
tokenizer = AutoTokenizer.from_pretrained("izumi-lab/deberta-v2-base-japanese")
|
34 |
+
model = AutoModelForMaskedLM.from_pretrained("izumi-lab/deberta-v2-base-japanese")
|
35 |
+
...
|
36 |
+
```
|
37 |
+
|
38 |
+
|
39 |
+
## Tokenization
|
40 |
+
|
41 |
+
The model uses a sentencepiece-based tokenizer, the vocabulary was trained on the Japanese Wikipedia using [sentencepiece](https://github.com/google/sentencepiece).
|
42 |
+
|
43 |
+
|
44 |
+
## Training Data
|
45 |
+
|
46 |
+
We used the following corpora for pre-training:
|
47 |
+
|
48 |
+
- [Japanese portion of CC-100](https://huggingface.co/datasets/izumi-lab/cc100-ja)
|
49 |
+
- [Japanese portion of mC4](https://huggingface.co/datasets/izumi-lab/mc4-ja-filter-ja-normal)
|
50 |
+
- [Japanese portion of OSCAR2301](izumi-lab/oscar2301-ja-filter-ja-normal)
|
51 |
+
- [Japanese Wikipedia as of July 20, 2023](https://huggingface.co/datasets/izumi-lab/wikipedia-ja-20230720)
|
52 |
+
- [Japanese Wikinews as of July 28, 2023](https://huggingface.co/datasets/izumi-lab/wikinews-ja-20230728)
|
53 |
+
|
54 |
+
|
55 |
+
We pretrained with the corpora mentioned above for 900k steps, and additionally pretrained with the following financial corpora for 100k steps:
|
56 |
+
- Summaries of financial results from October 9, 2012, to December 31, 2022
|
57 |
+
- Securities reports from February 8, 2018, to December 31, 2022
|
58 |
+
- News articles
|
59 |
+
|
60 |
+
|
61 |
+
## Training Parameters
|
62 |
+
|
63 |
+
learning_rate in parentheses indicate the learning rate for additional pre-training with the financial corpus.
|
64 |
+
- learning_rate: 2.4e-4 (6e-5)
|
65 |
+
- total_train_batch_size: 2,016
|
66 |
+
- max_seq_length: 512
|
67 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-06
|
68 |
+
- lr_scheduler_type: linear schedule with warmup
|
69 |
+
- training_steps: 1,000,000
|
70 |
+
- warmup_steps: 100,000
|
71 |
+
- precision: FP16
|
72 |
+
|
73 |
+
|
74 |
+
## Fine-tuning on General NLU tasks
|
75 |
+
|
76 |
+
We evaluate our model with the average of five seeds.
|
77 |
+
Other models are from [JGLUE repository](https://github.com/yahoojapan/JGLUE)
|
78 |
+
|
79 |
+
|
80 |
+
| Model | JSTS | JNLI | JCommonsenseQA |
|
81 |
+
|-------------------------------|------------------|-----------|----------------|
|
82 |
+
| | Pearson/Spearman | acc | acc |
|
83 |
+
| **DeBERTaV2 base** | **0.890/0.846** | **0.xxx** | **0.859** |
|
84 |
+
| Waseda RoBERTa base | 0.913/0.873 | 0.895 | 0.840 |
|
85 |
+
| Tohoku BERT base | 0.909/0.868 | 0.899 | 0.808 |
|
86 |
+
|
87 |
+
|
88 |
+
## Citation
|
89 |
+
|
90 |
+
TBA
|
91 |
+
|
92 |
+
|
93 |
+
## Licenses
|
94 |
+
|
95 |
+
The pretrained models are distributed under the terms of the [Creative Commons Attribution-ShareAlike 4.0](https://creativecommons.org/licenses/by-sa/4.0/).
|
96 |
+
|
97 |
+
|
98 |
+
## Acknowledgments
|
99 |
+
|
100 |
+
This work was supported in part by JSPS KAKENHI Grant Number JP21K12010, and the JST-Mirai Program Grant Number JPMJMI20B1, Japan.
|
config.json
ADDED
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": [
|
3 |
+
"DebertaV2ForMaskedLM"
|
4 |
+
],
|
5 |
+
"attention_probs_dropout_prob": 0.1,
|
6 |
+
"hidden_act": "gelu",
|
7 |
+
"hidden_dropout_prob": 0.1,
|
8 |
+
"hidden_size": 768,
|
9 |
+
"initializer_range": 0.02,
|
10 |
+
"intermediate_size": 3072,
|
11 |
+
"layer_norm_eps": 1e-07,
|
12 |
+
"max_position_embeddings": 512,
|
13 |
+
"max_relative_positions": -1,
|
14 |
+
"model_type": "deberta-v2",
|
15 |
+
"norm_rel_ebd": "layer_norm",
|
16 |
+
"num_attention_heads": 12,
|
17 |
+
"num_hidden_layers": 12,
|
18 |
+
"pad_token_id": 1,
|
19 |
+
"pooler_dropout": 0,
|
20 |
+
"pooler_hidden_act": "gelu",
|
21 |
+
"pooler_hidden_size": 768,
|
22 |
+
"pos_att_type": "p2c|c2p",
|
23 |
+
"position_biased_input": false,
|
24 |
+
"position_buckets": 256,
|
25 |
+
"relative_attention": true,
|
26 |
+
"share_att_key": true,
|
27 |
+
"torch_dtype": "float16",
|
28 |
+
"transformers_version": "4.31.0",
|
29 |
+
"type_vocab_size": 0,
|
30 |
+
"vocab_size": 32000
|
31 |
+
}
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:596337893657556b383e9813945f6cb3a990f5e0e90530d64b1d49941cd2ca37
|
3 |
+
size 542676485
|
special_tokens_map.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"bos_token": "[CLS]",
|
3 |
+
"cls_token": "[CLS]",
|
4 |
+
"eos_token": "[SEP]",
|
5 |
+
"mask_token": "[MASK]",
|
6 |
+
"pad_token": "[PAD]",
|
7 |
+
"sep_token": "[SEP]",
|
8 |
+
"unk_token": "[UNK]"
|
9 |
+
}
|
spm.model
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f8e9cbe24bc1bb25ef87a4371c222666539011d1a749cd4858a88a64771acc1a
|
3 |
+
size 804800
|
tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
tokenizer_config.json
ADDED
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"bos_token": "[CLS]",
|
3 |
+
"clean_up_tokenization_spaces": true,
|
4 |
+
"cls_token": "[CLS]",
|
5 |
+
"do_lower_case": false,
|
6 |
+
"eos_token": "[SEP]",
|
7 |
+
"mask_token": "[MASK]",
|
8 |
+
"model_max_length": 1000000000000000019884624838656,
|
9 |
+
"pad_token": "[PAD]",
|
10 |
+
"sep_token": "[SEP]",
|
11 |
+
"sp_model_kwargs": {},
|
12 |
+
"split_by_punct": false,
|
13 |
+
"tokenizer_class": "DebertaV2Tokenizer",
|
14 |
+
"unk_token": "[UNK]"
|
15 |
+
}
|