关于模型训练的几个问题？

1. 训练数据集可否不是连续的视频帧数据，而是想训练类似于 Coco 或者 Voc2017 的数据集？具体怎么配置呢？
2. 如何训练tiny和small模型？训练后的大小是否都应该是100多兆？需要怎么配置呢？
修改配置文件 sam2.1_hiera_b+_MOSE_finetune.yaml 的 checkpoint_path 的值为 /root/Code/Sam2-Model/sam2.1_hiera_tiny.pt ，但是训练后的模型大小仍然是800多兆。

Start by reviewing sam2.1_hiera_b+_MOSE_finetune.yaml and the training configuration for dataset input and checkpoint selection. Compare the COCO/VOC2017 question with the repository's supported training data, then investigate why using sam2.1_hiera_tiny.pt still produces an 800+ MB model. Done means the supported dataset setup, tiny/small configuration, and expected output sizes are documented.

Written by the indexing model from the issue text.