metadata

language: zh
widget:
  - text: '[CLS] 万 叠 春 山 积 雨 晴 ，'
  - text: '[CLS] 青 山 削 芙 蓉 ，'

Chinese GPT2 Language Models

Model description

This is the set of two Chinese GPT2 language models pre-trained by UER-py.

You can download the two Chinese GPT2 language models via HuggingFace from the links below:

Model	gpt2-chinese-poem	gpt2-chinese-couplet
Training data	Contains about 800,000 chinese ancient poems	Contains about 700,000 chinese couplets

How to use

Because the parameter skip_special_tokens is used in the pipelines.py , special tokens such as [SEP], [UNK] will be deleted, and the output results may not be neat.

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

When the parameter skip_special_tokens is True:

>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> from transformers import TextGenerationPipeline, 
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-poem")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-poem")
>>> text_generator = TextGenerationPipeline(model, tokenizer)   
>>> text_generator("[CLS]梅 山 如 积 翠 ，", max_length=50, do_sample=True)
    [{'generated_text': '[CLS]梅 山 如 积 翠 ， 的 手 堪 捧 。 遥 遥 仙 人 尉 ， 盘 盘 故 时 陇 。 丹 泉 清 可 鉴 ， 石 乳 甘 于 。 行 将 解 尘 缨 ， 于 焉 蹈 高 踵 。 我'}]

When the parameter skip_special_tokens is Flase:

>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> from transformers import TextGenerationPipeline, 
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-poem")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-poem")
>>> text_generator = TextGenerationPipeline(model, tokenizer)   
>>> text_generator("[CLS]梅 山 如 积 翠 ，", max_length=50, do_sample=True)
    [{'generated_text': '[CLS]梅 山 如 积 翠 ， 的 [UNK] 手 堪 捧 。 遥 遥 仙 人 尉 ， 盘 盘 故 时 陇 。 丹 泉 清 可 鉴 ， 石 乳 甘 可 捧 。 银 汉 迟 不 来 ， 槎 头 欲 谁 揽 。 何'}]

Training data

Contains about 800,000 chinese ancient poems.

Training procedure

Models are pre-trained by UER-py on Tencent Cloud TI-ONE. We pre-train 200,000 steps with a sequence length of 128.

python3 preprocess.py --corpus_path corpora/poem.txt \
                      --vocab_path models/google_zh_vocab.txt \  
                      --dataset_path poem.pt --processes_num 16 \
                      --seq_length 128 --target lm

python3 pretrain.py --dataset_path poem.pt \
                    --vocab_path models/google_zh_vocab.txt \
                    --output_model_path models/poem_gpt_base_model.bin \  
                    --config_path models/bert_base_config.json --learning_rate 5e-4 \
                    --tie_weight --world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
                    --batch_size 64 --report_steps 1000 \
                    --save_checkpoint_steps 50000 --total_steps 200000 \
                    --embedding gpt --encoder gpt2 --target lm

BibTeX entry and citation info

@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}
}