multi-image training in GRPO #354
Description
During multi-image training in GRPO, I adjusted the model in the LORA section. When using both accuracy and format as rewards, the model failed to learn any content—it only learned the output format. Is there a solution to this problem?