syspider

生成的时候可以使用past_key_values机制加速。博主生成的时候每一次输入都是整个序列，改进后能快不少

他复用了GPT2LMHeadModel这个类里的loss，这里面计算loss的时候已经错位了，你可以看源码

> Yes, the training data is under preparation. We will release it soon. quite excited

All the samples seem to have the same problem when the data source is "twitter_post"

> Hi, you are right and with that example code, the system message and user' instruction are used to calculate the loss and are learned by the model. Please refer...

> Hi, I agree with on that we could add some notes to the code, which we will. > > The thing is that most of time we find the...