Model predictions for each timestep

Need trained model files for all sequence lengths (like the original paper)

OR

One model trained with padded sequences

The first is probably fine, loading the models from disk doesn't take that long

No files, tests, or entry points are named. Start by clarifying whether the project should load separate trained models for each sequence length or use one model with padded sequences. Done means implementing the selected approach and providing model files or training support for all required sequence lengths.

Written by the indexing model from the issue text.