search

turboLJY / Transfer-Prompts-for-Text-Generation

16 stars 1 forks source link

Results in Daily Dialogue #8

Open gaojingsheng opened 2 years ago

gaojingsheng commented 2 years ago

Hi, I reproduce some of your experiment. The BART in personachat is epoch: 4 Bleu1: 44.64 Bleu2: 29.05 Distinct 1: 1.24 Distinct 2: 6.9, whose BLEU is higher than your results. And the more important is, all results in DailyDialogue exceed your results a lot. For example, the BART in DailyDialogue is epoch: 42 Bleu1: 64.0 Bleu2: 50.47 Distinct1: 5.57 Distinct2: 32.47. Also, the results of PTG in DailyDialogue are also better than your results(I have not run it over yet). Can you explain why I can run so many epochs in DailyDialogue and achieve total different results? If you have any questions about my reproduction, you can get in touch with me personally.

minji-o-j commented 1 year ago

Hello, I'm a researcher interested in this paper.

Is it the result of reproducing it by setting it to self.model.requires_grad_(True) (i.e., fine-tuning the BART model)?