Difference between best checkpoint performance and report in paper

[3i-cao]

Hello the author, I see on the provided best checkpoint: [best] epoch: 8 [loss] loss: 9.80451 [loss] ref_loss: 0.76449 [loss] lang_loss: 0.12697 [loss] lang_acc: 0.97148 [sco.] ref_acc: 0.45403 [sco.] iou_rate_0.25: 0.38326, iou_rate_0.5: 0.31289

and you paper report: IntanceRefer (Ours) Geo + RGB 77.45 66.83 31.27 24.77 40.23 32.93

What cause this performance gap? How to produce the results as the paper? Thank you so much!