@Wanfq I tried running a couple of different times, even with different tokenizer because I thought that was the issue, but after just one iteration, not even epoch, the model starts outputing gibberish. Have you guys seen this happened, and can only finetune before merging?
