This is a Japanese character-level GPT-2 Large (717M parameters) language model pre-trained on Japanese Wikipedia, the Japanese portion of CC-100, and the Japanese portion of OSCAR.

How to use

You can use this model directly with a pipeline for text generation.

python

>>> from transformers import pipeline, set_seed
>>> generator = pipeline('text-generation', model='ku-nlp/gpt2-large-japanese-char')
>>> set_seed(5)
>>> generator("<s>昨日私は京都で", max_length=30, do_sample=True, num_return_sequences=5)

[{'generated_text': '<s>昨日私は京都で仕事だったのですが、帰りは車を信号で止めて、'},
 {'generated_text': '<s>昨日私は京都で開かれた大阪市都市戦略会議に出席しました。そ'},
 {'generated_text': '<s>昨日私は京都で行われました関西の教育者・学校事例が集まるイ'},
 {'generated_text': '<s>昨日私は京都では初雪を見ました。朝は少しパッとしない天気で'},
 {'generated_text': '<s>昨日私は京都でこみっくトレジャーさんの撮影を見学させていた'}]

You can also use this model to get the features of a given text.

Vocabulary

A character-level vocabulary of size 6K is used. To be precise, rare characters may be split into bytes because byte-level byte-pair encoding (BPE) is used. The BPE tokenizer was trained on a small subset of the training data. Since the data were converted into a one-character-per-line format, merge operations never go beyond character boundaries.

Note that the tokenizer maps U+0020 to [UNK] because preprocessing eliminated whitespace characters (U+0020) from training data. Use U+3000 (Ideographic Space) instead.

Training data

We used the following corpora for pre-training:

Japanese Wikipedia (as of 20221020, 3.2GB, 27M sentences, 1.3M documents)
Japanese portion of CC-100 (85GB, 619M sentences, 66M documents)
Japanese portion of OSCAR (54GB, 326M sentences, 25M documents)

Note that we filtered out documents annotated with "header", "footer", or "noisy" tags in OSCAR. 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.

Training procedure

The training took about 8 months (with 7 interruptions) with a single NVIDIA A100 80GB GPU.

The following hyperparameters were used during pre-training:

learning_rate: 2e-4
per_device_train_batch_size: 6
gradient_accumulation_steps: 98
optimizer: AdamW with betas=(0.9, 0.999) and epsilon=1e-06
weight_decay: 0.01
lr_scheduler_type: linear
max_grad_norm: 1.0
max_steps: 500,000 (but terminated at 186,000 steps ~= 2.0 epochs)
warmup_steps: 10,000

The eval loss was 1.309 while the eval accuracy was 0.6890. The evaluation set consists of 5,000 randomly sampled documents from each of the training corpora.