I was trying to fine-tune the qwen-vl2-7b instruct model on a custom dataset. For that, I wrote the following collate function: @dataclass class Collator: processor: ProcessorMixin def __call__(self, ...

The mistake I was making was using two separate t5 text boxes, which is what I had hooked up to NAG. Once I started to increase the cfg, I needed to switch to the textencode node with both positive ...