Accumulation steps is 9 in the code

[taolusi]

As reported in the paper, the batch size for Cross-Document Coreference Resolution is 128, so the accumulation steps should be 8 when config.bathc_size is 16. But the accumulation steps seems to be 9 as in the code:

CDLM/cross_encoder/train.py

Line 164 in 8e7a844

if (i + 1) % 9== 0: