# D-JEPA: pusht-relational

- Architecture: `set_aligner`.
- Tensor count: 30.
- Includes pretrained predictor weights: false.
- Original checkpoint SHA256: `84c2b0fb93afdad9279b220ecf696642528899bb233a21ab1489c31ec7359d9f`.
- Exported tensor-file SHA256: `efea4301f605f681c7f85f934a6ecdcafa9984058f72be3be2d625bfc9e16941`.

Load `model.pt` with `torch.load(..., map_location="cpu", weights_only=True)`. Use the matching D-JEPA code profile and JSON configuration. Tensor values are unchanged; metadata and optimizer state have been separated.

Input width: 386. Gate: `base_minus_refined`, strict threshold `-0.03162526342176623`; quantization decimals: `None`. Candidate IDs and feature preprocessing must match the task profile.

Project license is selected by the authors before public upload.