training on multiple datasets

Hello, we have encountered some issues when training on multiple datasets and would like to ask for your advice. During the training of SAM3, a large amount of data is also used. Should all datasets be merged into a single COCO-format JSON annotation file for one-time training, or is it feasible to train on different datasets sequentially? If sequential training is adopted, are there any recommended strategies—such as mixing in a portion of previously used data—to avoid catastrophic forgetting of earlier learned knowledge?

Review the repository's SAM3 finetuning documentation and example notebooks first. Determine whether multiple COCO-format datasets can be combined or trained sequentially, and whether any documented strategy addresses forgetting. Done should be a clear supported workflow or a documented limitation with recommended guidance.