Hi,
While testing ModelDataSet with a sliding window (DatasetConfig(window_size=N)), I ran into a crash once the window limit is reached.

In DataRecorder._store_dataset_snapshot, eviction assumes blocks are stored as (time, data) tuples:

_old_time, old_data = storage.blocks[0]

However, for modeldataset, blocks are stored as a single merged dict:

row = {**data, "time": time}
storage.blocks.append(row)

So when eviction tries to unpack storage.blocks[0], it ends up unpacking a dict and raises:

ValueError: too many values to unpack

This only happens when:

• Using ModelDataSet

• Setting window_size

• Running long enough to trigger eviction

It looks like an internal inconsistency — other dataset types store (time, data), but modeldataset stores a dict directly.
A minimal fix would be to store modeldataset blocks as (time, data) like the others, and handle merging "time" during DataFrame conversion instead.

I can open a small PR if this approach is correct.