Closed holylovenia closed 1 year ago
By the way, I have uncommented the comments and also updated the requested results in the sheets. Also, during the inspection, I also found that running the fine-tuning with fp16 would cause nan
loss.
@holylovenia : Thank you for inspecting the code and the experiment. We can discuss the result further on the next meeting. Thank you!
For details, see commit messages. Performances have been reported in the sheets for comparison.