Inquiry About Reproducing VIM_Tiny Results

[lvjindi]

I greatly appreciate your work on VIM and find it very inspiring. I am currently building upon your research but encountered some difficulties in reproducing the results for vim_tiny. Specifically, when following the settings provided in .\Vim\vim\scripts\ft-vim-t.sh, I am unable to achieve the reported 76.1 accuracy and can only reach 74.x.

The only notable difference in my setup is that I am using 4 GPUs with a gradient accumulation step of 2. I would be grateful for any insights you could share—are there any additional tricks or implementation details that might help achieve the reported results?