HoloD3-models

HoloD3 model artifacts

This private model repository stores the exact PyTorch checkpoints used by the HoloD3 reproduction package. The source code, annotations, training inputs, and artifact manifests live in the private GitHub repository. Model paths in this repository intentionally match their checkout locations under models/.

Download

First obtain read access to this private model repository. Authenticate without placing a token in a command-line argument:

uv run hf auth login

From a HoloD3 checkout, download only the four operational checkpoints:

uv run holod3 fetch-models --scope production
uv run holod3 verify

Download the production checkpoints plus the exact YOLO training initializers:

uv run holod3 fetch-models --scope reproduction

Download all 17 archived production, initializer, baseline, retrained, and comparison checkpoints:

uv run holod3 fetch-models --scope all

scripts/fetch_models.py exposes the same operation for users who prefer a script entry point. Every downloaded file is checked against its recorded byte size and SHA-256 digest in models/remote_manifest.json.

Production chain

The production rules, thresholds, provenance, and hashes are recorded in models/production/manifest.json. See the GitHub repository's docs/model-card.md, docs/parity.md, and docs/training.md for the intended domain, limitations, exact fallback rules, and complete retraining procedure.

Access and licensing

This repository is private because no public redistribution license has been selected for the project-specific checkpoints or training data. The production YOLO checkpoint contains Ultralytics AGPL-3.0 metadata. Access does not grant permission to redistribute artifacts; consult the repository owner and the notices in the source repository before deployment or redistribution.