How to reproduce results on the summarization task? about fairseq HOT 3 CLOSED

We used the data from here: https://github.com/facebookarchive/NAMAS
The remaining details are in the paper, e.g., vocab of 30K, hidden size 256, batchsize 128 etc.

from fairseq.

Hi,
I read the paper that additionally you require outputs to be at least 14 words long, but I think length of much target summarizations may less than 14. And I use the training parameters as following:
--lr 0.25 --clip-norm 0.1 --dropout 0.1 --max-tokens 4000 --arch fconv_giga
The architecture of fconv_giga is the same as the paper said.
May I ask for some suggestion?
Thank you very much!

from fairseq.

This looks good. You may need to tune the learning rate and --max-tokens but other than that this looks right. Since publishing the paper we changed the batching strategy, so you may have to experiment in getting a good setting for this problem.

from fairseq.