Can't reproduce result.

[piaoyun2022]

Hi, I'm recently try to reproduce your result. I use Deepseek-V3 as the base model. It seems that raw V3 have 41.89% EM on total CWQ test dataset, and Think-on-Graph somewhat gives 45%.
Could you give some advice about that? Thanks!

[Atroxgod]

Hi, I'm recently try to reproduce your result. I use Deepseek-V3 as the base model. It seems that raw V3 have 41.89% EM on total CWQ test dataset, and Think-on-Graph somewhat gives 45%. Could you give some advice about that? Thanks!

I found the CWQ train dataset and saw that each question has quite a few answers (with the same meaning, aliases), yet the author's CWQ dev dataset only provides one answer. How did you achieve a benchmark of 40?