nobu-g commited on
Commit
8298794
1 Parent(s): 7cd828b

first commit

Browse files
Files changed (6) hide show
  1. README.md +86 -0
  2. config.json +37 -0
  3. pytorch_model.bin +3 -0
  4. special_tokens_map.json +7 -0
  5. tokenizer_config.json +19 -0
  6. vocab.txt +0 -0
README.md ADDED
@@ -0,0 +1,86 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ - character
10
+ - wwm
11
+ datasets:
12
+ - wikipedia
13
+ - cc100
14
+ - oscar
15
+ metrics:
16
+ - accuracy
17
+ mask_token: "[MASK]"
18
+ widget:
19
+ - text: "京都大学で自然言語処理を[MASK][MASK]する。"
20
+ ---
21
+
22
+ # Model Card for Japanese character-level DeBERTa V2 base
23
+
24
+ ## Model description
25
+
26
+ 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.
27
+ This model is trained with character-level tokenization and whole word masking.
28
+
29
+ ## How to use
30
+
31
+ You can use this model for masked language modeling as follows:
32
+
33
+ ```python
34
+ from transformers import AutoTokenizer, AutoModelForMaskedLM
35
+ tokenizer = AutoTokenizer.from_pretrained('ku-nlp/deberta-v2-base-japanese-char-wwm')
36
+ model = AutoModelForMaskedLM.from_pretrained('ku-nlp/deberta-v2-base-japanese-char-wwm')
37
+
38
+ sentence = '京都大学で自然言語処理を[MASK][MASK]する。'
39
+ encoding = tokenizer(sentence, return_tensors='pt')
40
+ ...
41
+ ```
42
+
43
+ You can also fine-tune this model on downstream tasks.
44
+
45
+ ## Tokenization
46
+
47
+ There is no need to tokenize texts in advance, and you can give raw texts to the tokenizer.
48
+ The texts are tokenized into character-level tokens by [sentencepiece](https://github.com/google/sentencepiece).
49
+
50
+ ## Training data
51
+
52
+ We used the following corpora for pre-training:
53
+
54
+ - Japanese Wikipedia (as of 20221020, 3.2GB, 27M sentences, 1.3M documents)
55
+ - Japanese portion of CC-100 (85GB, 619M sentences, 66M documents)
56
+ - Japanese portion of OSCAR (54GB, 326M sentences, 25M documents)
57
+
58
+ Note that we filtered out documents annotated with "header", "footer", or "noisy" tags in OSCAR.
59
+ 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.
60
+
61
+ ## Training procedure
62
+
63
+ We first segmented texts in the corpora into words using [Juman++ 2.0.0-rc3](https://github.com/ku-nlp/jumanpp/releases/tag/v2.0.0-rc3) for whole word masking.
64
+ Then, we built a sentencepiece model with 32000 tokens including all characters that appear in the training corpus.
65
+
66
+ We tokenized raw corpora into character-level subwords using the sentencepiece model and trained the Japanese DeBERTa model using [transformers](https://github.com/huggingface/transformers) library.
67
+ The training took 20 days using 8 NVIDIA A100-SXM4-40GB GPUs.
68
+
69
+ The following hyperparameters were used during pre-training:
70
+
71
+ - learning_rate: 2e-4
72
+ - per_device_train_batch_size: 46
73
+ - distributed_type: multi-GPU
74
+ - num_devices: 8
75
+ - gradient_accumulation_steps: 6
76
+ - total_train_batch_size: 2,208
77
+ - max_seq_length: 512
78
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-06
79
+ - lr_scheduler_type: linear schedule with warmup (lr = 0 at 500k steps)
80
+ - training_steps: 320,000
81
+ - warmup_steps: 10,000
82
+
83
+ ## Acknowledgments
84
+
85
+ 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".
86
+ 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": "4.25.1",
35
+ "type_vocab_size": 0,
36
+ "vocab_size": 22012
37
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c48bf25170e9a7e91a2e25b927d6c7a34470b9c5859fdc9d5e8940bfb4003cbc
3
+ size 486750941
special_tokens_map.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "cls_token": "[CLS]",
3
+ "mask_token": "[MASK]",
4
+ "pad_token": "[PAD]",
5
+ "sep_token": "[SEP]",
6
+ "unk_token": "[UNK]"
7
+ }
tokenizer_config.json ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "cls_token": "[CLS]",
3
+ "do_lower_case": false,
4
+ "do_subword_tokenize": true,
5
+ "do_word_tokenize": true,
6
+ "jumanpp_kwargs": null,
7
+ "mask_token": "[MASK]",
8
+ "mecab_kwargs": null,
9
+ "model_max_length": 1000000000000000019884624838656,
10
+ "never_split": null,
11
+ "pad_token": "[PAD]",
12
+ "sep_token": "[SEP]",
13
+ "special_tokens_map_file": null,
14
+ "subword_tokenizer_type": "character",
15
+ "sudachi_kwargs": null,
16
+ "tokenizer_class": "BertJapaneseTokenizer",
17
+ "unk_token": "[UNK]",
18
+ "word_tokenizer_type": "basic"
19
+ }
vocab.txt ADDED
The diff for this file is too large to render. See raw diff