Update README.md

README.md

@@ -1,3 +1,90 @@
 ---
+language: ja
 license: cc-by-sa-4.0
+library_name: transformers
+tags:
+  - roberta
+  - fill-mask
+datasets:
+  - wikipedia
+  - cc100
+mask_token: "[MASK]"
+widget:
+  - text: "京都 大学 で [MASK] を 専攻 する 。"
+  - text: "東京 は 日本 の [MASK] だ 。"
+  - text: "カフェ で [MASK] を 注文 する 。"
+  - text: "[MASK] 名人 が タイトル の 防衛 に 成功 する 。"
 ---
+
+# ku-accms/roberta-base-japanese-ssuw
+## Model description
+This is a pre-trained Japanese RoBERTa base model for super short unit words (SSUW).
+
+## Pre-processing
+The input text should be converted to full-width (zenkaku) characters and segmented into super short unit words in advance (e.g., by KyTea).
+
+## How to use
+You can use this model directly with a pipeline for masked language modeling:
+
+```python
+>>> from transformers import pipeline
+>>> unmasker = pipeline('fill-mask', model='ku-accms/roberta-base-japanese-ssuw')
+>>> unmasker("京都 大学 で [MASK] を 専攻 する 。")
+[{'sequence': '京都 大学 で 文学 を 専攻 する 。',
+  'score': '0.1479644924402237',
+  'token': '17907',
+  'token_str': '文学'}
+ {'sequence': '京都 大学 で 哲学 を 専攻 する 。',
+  'score': '0.07658644765615463',
+  'token': '19302',
+  'token_str': '哲学'}
+ {'sequence': '京都 大学 で デザイン を 専攻 する 。',
+  'score': '0.06302948296070099',
+  'token': '14411',
+  'token_str': 'デザイン'}
+ {'sequence': '京都 大学 で 建築 を 専攻 する 。',
+  'score': '0.060596249997615814',
+  'token': '15478',
+  'token_str': '建築'}
+ {'sequence': '京都 大学 で 工学 を 専攻 する 。',
+  'score': '0.0574776753783226',
+  'token': '18632',
+  'token_str': '工学'}
+```
+
+Here is how to use this model to get the features of a given text in PyTorch:
+```python
+import zenhan
+import Mykytea
+kytea_model_path = "somewhere"
+kytea = Mykytea.Mykytea("-model {} -notags".format(kytea_model_path))
+def preprocess(text):
+    return " ".join(kytea.getWS(zenhan.h2z(text)))
+
+from transformers import BertTokenizer, RobertaModel
+tokenizer = BertTokenizer.from_pretrained('ku-accms/roberta-base-japanese-ssuw')
+model = BertModel.from_pretrained("ku-accms/roberta-base-japanese-ssuw")
+text = "京都大学で自然言語処理を専攻する。"
+encoded_input = tokenizer(preprocess(text), return_tensors='pt')
+output = model(**encoded_input)
+```
+
+## Training data
+We used a Japanese Wikipedia dump (as of 20230101, 3.3GB) and a Japanese portion of CC100 (70GB).
+
+## Training procedure
+We first segmented the texts into words by KyTea and then tokenized the words into subwords using WordPiece with a vocabulary size of 32,000. We pre-trained the BERT model using [transformers](https://github.com/huggingface/transformers) library. The training took about 7 days using 4 NVIDIA A100-SXM4-80GB GPUs.
+
+The following hyperparameters were used for the pre-training.
+
+- learning_rate: 1e-4
+- weight decay: 1e-2
+- per_device_train_batch_size: 80
+- num_devices: 4
+- gradient_accumulation_steps: 3
+- total_train_batch_size: 960
+- max_seq_length: 512
+- optimizer: AdamW with betas=(0.9,0.999) and epsilon=1e-06
+- lr_scheduler_type: linear schedule with warmup
+- training_steps: 500,000
+- warmup_steps: 10,000