D-JEPA checkpoints

Project website · Model repository · Decision-supervision dataset · Code

Task/module profiles for D-JEPA: A Decision-Aligned Latent World Model. See each profile's README.md and config.json for architecture, provenance, upstream dependencies, calibrated decisions and tensor-file checksums.

Choose a checkpoint profile

Task	Profile	Contents
PushT	Relational alignment	Learned decision-alignment module
PushT	Predictive adaptation	Updated parameters; requires the exact base model
PushT	Adapted predictor	Full predictor weights
PushT	Multi-geometry alignment	Module using four predictive geometries
PushT	Exact representation realization	Two predictors and relation module in one tensor file
Reacher	Relational alignment	Task-specific decision-alignment module
Reacher	Temporal transport	Bounded future-update module
Reacher	Full model	Predictors, alignment and transport weights
Granular	Relational alignment	Spatial-feature alignment module
Granular	Multi-view alignment	Four-view variant using the same backbone
PushObj	Unseen-shape transfer	Task-local ordinal alignment module
PushT	Visual-shift alignment	Corruption-trained ordinal alignment module

These are configurations of D-JEPA, not 12 separate methods. Each profile contains model.pt, config.json and its own card. The manifest indexes source and exported weight hashes.

Loading and dependencies

Download an individual checkpoint profile with the Hugging Face CLI:

hf download Shuaijun/D-JEPA \
  pusht-relational/model.pt pusht-relational/config.json pusht-relational/README.md \
  --local-dir checkpoints/D-JEPA

This repository hosts custom PyTorch checkpoint profiles, not dataset splits. It does not require the Dataset Viewer. The linked supervision dataset is a downloadable ZIP/NPZ archive; see its card for extraction and loading instructions.

All model.pt files contain tensor dictionaries and support torch.load(path, map_location="cpu", weights_only=True). Exact tensor values from selected formal checkpoints are preserved. Packaging removes optimizer state and private path metadata; exported file hashes are therefore distinct from original training-file hashes.

import torch

state = torch.load("checkpoints/D-JEPA/pusht-relational/model.pt",
                   map_location="cpu", weights_only=True)

This reads a tensor state dictionary, not an instantiated model. Use the matching D-JEPA architecture and configuration; these custom modules are not loaded through Transformers AutoModel.from_pretrained.

Profiles with upstream_included: true contain pretrained predictors. Other profiles contain our learned modules or updated parameter subsets and require their recorded predictive models/features. pusht-exact-realization-full is one file containing two predictors and a relational operator, not a distilled single-backbone student.

Reacher's physical result uses the relational selector; temporal transport is also provided but its native-distance diagnostic is not a substitute result. Multi-geometry's 128-start development result remains a separate protocol from the independent 256-start PushT evaluation.

Local release candidate: authors have not yet selected publication licenses or completed all upstream weight redistribution checks. This card grants no new rights to third-party material. The author-designated repository is Shuaijun/D-JEPA.