openai的gpt-2模型最近在风口浪尖上。Language Models are Unsupervised Multitask Learners论文已经出来,但是由于该模型没有将训练过程开源出来,所以本博客仅仅是针对已经公布的117M的预训练模型进行测试。
1、论文贡献
In this paper, we connect these two lines of work and continue the trend of more general methods of transfer. We demonstrate language models can perform down-stream tasks in a zero-shot setting – without any parameter or architecture modification. We demonstrate this approach shows potential by highlighting the ability of language models to perform a wide range of tasks in a zero-shot setting. We achieve promising, competitive, and state of the art results depending on the task.
找更大数量的无监督训练数据来执行多任务学习,使模型更具泛化能力。论文实验也证明了该模型具有惊人的效果。
该论文的模型大部分还是遵循GPT-1的模型,但有两点不同的是:
(1)训练数据集更加庞大;
(2)在第二阶段时候,无监督地做多样性的任务。
2、117M的实验测试
执行测试程序,效果如下:
其中任选一个例子,可以看到对话的自动生成效果,可读性还是非常好的。