forward_step_eval Error

I encountered an error while executing eval during the model training process.
[rank2]: objects_to_broadcast = [batch["txt"], broadcast_shape, batch["points"]] # 准备广播对象
[rank2]: KeyError: 'points'`

But I checked the data loaded in each batch, and it only contains 4 keys. This is defined in the **SFTDataset**：
`item = {
"mp4": tensor_frms,
"txt": caption,
"num_frames": num_frames,
"fps": fps,
}`

Start at the forward_step_eval entry point and compare its expected batch keys with the SFTDataset item definition shown in the issue. Trace where the batch is assembled and determine which interface is intended; done means evaluation no longer raises KeyError for points and the dataset/evaluation contract is consistent.

Written by the indexing model from the issue text.