Falcon -7B training loss not reducing

[pcvishak]

Thanks for the wonderful code. While training Falcon-7B on Alpaca dataset, the training loss is not reducing. It used to work fine. Were there any recent changes?