nobu-g commited on
Commit
11debab
1 Parent(s): a3d0f70

first commit

Browse files
README.md ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: ja
3
+ license: cc-by-sa-4.0
4
+ library_name: transformers
5
+ tags:
6
+ - deberta
7
+ - deberta-v2
8
+ - fill-mask
9
+ datasets:
10
+ - wikipedia
11
+ - cc100
12
+ - oscar
13
+ metrics:
14
+ - accuracy
15
+ mask_token: "[MASK]"
16
+ widget:
17
+ - text: "京都大学で自然言語処理を[MASK]する。"
18
+ ---
19
+
20
+ # Model Card for Japanese DeBERTa V2 base
21
+
22
+ ## Model description
23
+
24
+ This is a Japanese DeBERTa V2 base model pre-trained on Japanese Wikipedia, the Japanese portion of CC-100, and the Japanese portion of OSCAR.
25
+
26
+ ## How to use
27
+
28
+ You can use this model for masked language modeling as follows:
29
+
30
+ ```python
31
+ from transformers import AutoTokenizer, AutoModelForMaskedLM
32
+ tokenizer = AutoTokenizer.from_pretrained('ku-nlp/deberta-v2-base-japanese')
33
+ model = AutoModelForMaskedLM.from_pretrained('ku-nlp/deberta-v2-base-japanese')
34
+
35
+ sentence = '京都大学で自然言語処理を[MASK]する。'
36
+ encoding = tokenizer(sentence, return_tensors='pt')
37
+ ...
38
+ ```
39
+
40
+ You can also fine-tune this model on downstream tasks.
41
+
42
+ ## Tokenization
43
+
44
+ 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).
45
+
46
+ ## Training Data
47
+
48
+ We used the following corpora for pre-training:
49
+
50
+ - Japanese Wikipedia (as of 20221020, 3.2GB, 27M sentences, 1.3M documents)
51
+ - Japanese portion of CC-100 (85GB, 619M sentences, 66M documents)
52
+ - Japanese portion of OSCAR (54GB, 326M sentences, 25M documents)
53
+
54
+ Note that we filtered out documents annotated with "header", "footer", or "noisy" tags in OSCAR.
55
+ 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.
56
+
57
+ ## Training procedure
58
+
59
+ We first segmented texts in the corpora into words using [Juman++](https://github.com/ku-nlp/jumanpp).
60
+ 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).
61
+
62
+ 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.
63
+ The training took three weeks using 8 NVIDIA A100-SXM4-40GB GPUs.
64
+
65
+ The following hyperparameters were used during pre-training:
66
+
67
+ - learning_rate: 2e-4
68
+ - per_device_train_batch_size: 44
69
+ - distributed_type: multi-GPU
70
+ - num_devices: 8
71
+ - gradient_accumulation_steps: 6
72
+ - total_train_batch_size: 2,112
73
+ - max_seq_length: 512
74
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-06
75
+ - lr_scheduler_type: linear schedule with warmup
76
+ - training_steps: 500,000
77
+ - warmup_steps: 10,000
78
+
79
+ The accuracy of the trained model on the masked language modeling task was 0.779.
80
+ The evaluation set consists of 5,000 randomly sampled documents from each of the training corpora.
81
+
82
+ ## Fine-tuning on NLU tasks
83
+
84
+ <!-- https://github.com/yahoojapan/JGLUE -->
85
+ Coming soon.
86
+
87
+ ## Acknowledgments
88
+
89
+ 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".
90
+ For training models, we used the mdx: a platform for the data-driven future.
config.json ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "DebertaV2ForMaskedLM"
4
+ ],
5
+ "attention_head_size": 64,
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "conv_act": "gelu",
8
+ "conv_kernel_size": 3,
9
+ "hidden_act": "gelu",
10
+ "hidden_dropout_prob": 0.1,
11
+ "hidden_size": 768,
12
+ "initializer_range": 0.02,
13
+ "intermediate_size": 3072,
14
+ "layer_norm_eps": 1e-07,
15
+ "max_position_embeddings": 512,
16
+ "max_relative_positions": -1,
17
+ "model_type": "deberta-v2",
18
+ "norm_rel_ebd": "layer_norm",
19
+ "num_attention_heads": 12,
20
+ "num_hidden_layers": 12,
21
+ "pad_token_id": 0,
22
+ "pooler_dropout": 0,
23
+ "pooler_hidden_act": "gelu",
24
+ "pooler_hidden_size": 768,
25
+ "pos_att_type": [
26
+ "p2c",
27
+ "c2p"
28
+ ],
29
+ "position_biased_input": false,
30
+ "position_buckets": 256,
31
+ "relative_attention": true,
32
+ "share_att_key": true,
33
+ "torch_dtype": "float32",
34
+ "transformers_version": "4.23.1",
35
+ "type_vocab_size": 0,
36
+ "vocab_size": 32000
37
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:57fb86bf188376f5d1230e5835270f9e7bb00e5db5aaedfb9445df7b571ba9c3
3
+ size 548197213
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 sha256:6c111c16e2e52366dcac46b886e40650bb843fe2938a65f5970271fc5697a127
3
+ size 805061
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": "[CLS]",
3
+ "cls_token": "[CLS]",
4
+ "do_lower_case": false,
5
+ "eos_token": "[SEP]",
6
+ "keep_accents": true,
7
+ "mask_token": "[MASK]",
8
+ "name_or_path": "data/deberta_v2_tokenizer",
9
+ "pad_token": "[PAD]",
10
+ "sep_token": "[SEP]",
11
+ "sp_model_kwargs": {},
12
+ "special_tokens_map_file": null,
13
+ "split_by_punct": false,
14
+ "tokenizer_class": "DebertaV2Tokenizer",
15
+ "unk_token": "[UNK]"
16
+ }