can't get the result reported in paper when using end2end training without dbsearch

[fasterbuild]

Hi all,
with default parameters, I run end2end training without desearch, train file is resources/gpt2/train.history_belief_action_sys_delex with 56778 samples in the file. The ppl in valid set stop decreasing after only 2 epoch, got final valid set ppl=2.30. The success rate of test set is around 18.5%, much lower than 70.5% reported in paper table 3. And the belief acc is around 42%, also much lower than 55% in paper table 1. I wonder if the model is trained well with default params, would you please release your hyper parameters for end2end training and traing details?

[gungui98]

Hi @fasterbuild , I have the same problem as you and I don't even get the same accuracy as you did! could you share the hyper-params you are using?