ARGUS retained research baselines
Four own trained checkpoints, released as research baselines by Team Cinder Covenant. They use community CT and Villa / ScrollPrize dinovol architectures. These are not qualified ink detectors, new ink-accuracy results or unread-scroll readings. The four original files total 4,125,955,140 bytes. Check the individual cards for exact hashes, exposure and historical results.
- PHerc1203 MAE v4b: high-resolution reconstruction continuation.
- PHerc0268 MAE epoch29: random-init coarse reconstruction baseline.
- PHerc0268 small DINO/iBOT step3000: own proof-config training and continuation.
- Multiscroll coarse DINO/iBOT step3750: negative baseline with PHerc0009B pretraining exposure.
Verify a download without loading a model
The individual cards give direct download URLs after publication. Download the four exact files into the subdirectories named in MODEL_MANIFEST.json, then run:
python -B inspect_checkpoint.py --root . --manifest MODEL_MANIFEST.json
This standard-library helper checks file size and SHA-256, inventories Torch ZIP metadata and reads bounded pickle opcodes. It never calls pickle.load, torch.load or a model. It does not download automatically. Its metadata reports contain counts and key names, not embedded private path values.
Loading / reproduction scope
The original checkpoints retain training state, not just a slim inference backbone. Exact historical training environments have not been recovered for every file. Audit source pins and sanitized configurations are supplied; they are not proof of training-time source parity. A load/forward pass was not performed for this publication.
For Villa MAE, historical code reconstructs NetworkFromConfig using the checkpoint's embedded model_config and then loads model with strict=True; its ConfigWrapper carries model_config, train_patch_size, train_batch_size and spacing. For dinovol, the historical trainer loads separate student, teacher, optimizer, scaler and RNG state; inspect dinovol_2/pretrain.py at the card's pin. Do not assume the entire resume artifact is a single model state_dict, or that the teacher/student projection heads equal an ink detector.
For any future loader, use an explicitly CPU-only, no-forward, no-data-access check with exact dependencies. Prefer PyTorch's weights-only restricted loading if the archive supports it. If compatibility requires unsafe pickle deserialization or missing globals, stop and review the original trusted artifact and source; do not silently switch to unrestricted loading. This publication does not supply an unverified loader or automatically execute a historical target-evaluation/training script. Safetensors or slim teacher exports remain future work and are not represented as present.
The public PHerc0268 starter provides source coordinates and an explicit opt-in reconstruction recipe; downloaded training data still need their own terms and coordinate checks. The broader ARGUS research collection records corrected historical metrics and limitations.
Licenses
Owner-released weights: CC BY-NC 4.0, non-commercial with attribution. Cite Team Cinder Covenant (Ryan Gurganious and Daine Ball), the individual checkpoint, exact revision/hash, upstream architecture authors and Vesuvius Challenge CT source. Original CT terms continue separately. Code: upstream Villa MIT; dinovol root MIT and relevant Meta-derived dinovol v2 implementation Apache-2.0. Preserved notices are under LICENSES. The new metadata-only helper is MIT.
The original DINO files retain historical local paths and resumable training state. These are disclosed archival metadata, not instructions to access an author's workstation. Scoped metadata review found no credential patterns or explicit credential-key strings. File hashes establish identity, not decoded compatibility, scientific quality or model safety.