Chinese Ancient GPT2 Model

Model description

The model is pre-trained by UER-py, which is introduced in this paper. Besides, the model could also be pre-trained by TencentPretrain introduced in this paper, which inherits UER-py to support models with parameters above one billion, and extends it to a multimodal pre-training framework.

The model is used to generate ancient Chinese. You can download the model from the UER-py Modelzoo page, or GPT2-Chinese Github page, or via HuggingFace from the link gpt2-chinese-ancient

How to use

You can use the model directly with a pipeline for text generation:

>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-ancient")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-ancient")
>>> text_generator = TextGenerationPipeline(model, tokenizer)   
>>> text_generator("当是时", max_length=100, do_sample=True)
    [{'generated_text': '[CLS]当是时 所 议 者 不 为 无 据 , 况 亦 在 之 列 乎 ? 然 则 今 日 之 事 , 所 当 思 者 在 何 ? 欲 求 国 是 于 天 下 , 莫 在 于 得 人 。 臣 以 为 求 人 之 法 , 不 在 多 用 官 一 途 。 诚 使 得 才 者 众 , 人 才 者 优 , 则 治 所 当 得 , 而 不 事 于 官 者 , 人 才 乃 其 常 也 。 所 当 讲 者'}]

Training data

Training data contains 3,000,000 ancient Chinese which are collected by daizhigev20. Since part of ancient corpus has no punctuation, we used the ancient Chinese punctuation system developed by BNU ICIP lab. 

Training procedure

The model is pre-trained by UER-py on Tencent Cloud. We pre-train 500,000 steps with a sequence length of 320. We use extended vocabulary to handle out-of-vocabulary words. The Chinese character that occurs greater than or equal to 100 in ancient Chinese corpus is added to the vocabulary.

python3 preprocess.py --corpus_path corpora/ancient_chinese.txt \
                      --vocab_path models/google_zh_ancient_vocab.txt \
                      --dataset_path ancient_chinese_dataset.pt --processes_num 16 \
                      --seq_length 320 --data_processor lm
python3 pretrain.py --dataset_path ancient_chinese_dataset.pt \
                    --vocab_path models/google_zh_ancient_vocab.txt \
                    --config_path models/bert_base_config.json \
                    --output_model_path models/ancient_chinese_gpt2_model.bin \
                    --world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
                    --total_steps 500000 --save_checkpoint_steps 100000 --report_steps 10000 \
                    --learning_rate 5e-4 --batch_size 32

Finally, we convert the pre-trained model into Huggingface's format:

python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path models/ancient_chinese_gpt2_model.bin-500000 \
                                                        --output_model_path pytorch_model.bin \
                                                        --layers_num 12

BibTeX entry and citation info

@article{radford2019language,
  title={Language Models are Unsupervised Multitask Learners},
  author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
  year={2019}
}

@article{zhao2019uer,
  title={UER: An Open-Source Toolkit for Pre-training Models},
  author={Zhao, Zhe and Chen, Hui and Zhang, Jinbin and Zhao, Xin and Liu, Tao and Lu, Wei and Chen, Xi and Deng, Haotang and Ju, Qi and Du, Xiaoyong},
  journal={EMNLP-IJCNLP 2019},
  pages={241},
  year={2019}
}

@article{zhao2023tencentpretrain,
  title={TencentPretrain: A Scalable and Flexible Toolkit for Pre-training Models of Different Modalities},
  author={Zhao, Zhe and Li, Yudong and Hou, Cheng and Zhao, Jing and others},
  journal={ACL 2023},
  pages={217},
  year={2023}
}
Downloads last month
346
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Space using uer/gpt2-chinese-ancient 1