NingMiao

We actually trained a backward GPT-2 model. It is not difficult since the code for training is released. I also recommend using BERT or just a forward GPT-2 as the...

Please try to download 1-billion dataset at https://www.statmt.org/lm-benchmark/.

您好。 我也尝试过中文数据集，但是只在specific domain上做过，效果还可以呀。关键是中文的language model训好就可以了，强烈建议使用词而不是字作为基本的单元，这样就可以直接套原来的模型。

1.ZPar有貌似有中文模型，可以直接换用，如果没有，就需要按论文搞一个。 2.这个应该没什么问题，不过我们可以邮件具体交流。 3.论文中邮箱哦，你可以随时联系我。就不粘在这里了，以防各种垃圾邮件。

Thank you for pointing out the problem. The problem results from a broken checkpoint file and I have already updated the file. Some path problems have also been fixed. Please...

Since I'm not authorized to release 1-billion dataset in my code, please download it from its official website. By the way, you can use any dataset to train the language...