Datasets:
Studio ink model (PHerc0814 / PHerc0841 / PHerc0009B) — training data, weights and checkpoint
Everything needed to train the Studio ink model from random initialization, or to continue the paused run,
without access to the original machines or scans. Code and instructions:
GitHub priyankaranke/vesuvius-studio-multiscroll-cpt
(MIT; the same base training stack plus the later training stages). Published weights and model card:
HF paranke/vesuvius-814-841-9b-resunet-studio-20260926.
Download
The file paths mirror the code repository's layout, so the dataset downloads into the same working folder as the code:
git clone https://github.com/priyankaranke/vesuvius-studio-multiscroll-cpt.git work
hf download paranke/vesuvius-814-841-9b-studio-training-data --repo-type dataset --local-dir work
The frozen data.json records absolute /home/ubuntu/... paths; use /home/ubuntu as the working folder or symlink
/home/ubuntu/donor-studio-resunet-scratch-20260926 to work/donor-studio-resunet-scratch-20260926.
Contents
| Path | What |
|---|---|
donor-studio-resunet-scratch-20260926/prepared/<dataset>/ |
All 210 snapshot datasets have selection.json; the 202 eligible ones the plan and labels; the 199 with samples (198 training + the holdout) ct-tiles.npy and render.json (8,340 tiles, 32.6 GiB): labels.npy (full canvas, uint8: 0 unknown, 1 background, 2 ink), coordinates.npy, entry-to-patch.npy, tiles.npy (512 px tile origins), ct-tiles.npy ([tiles, 512, 512, 16] uint8 native CT, model depth order), selection.json, render.json |
donor-studio-resunet-scratch-20260926/labels.tar |
Frozen Donor Studio label snapshot (SQLite / TIFF / PNG, cutoff 2026-09-26 05:37 UTC) the plans were derived from; tar xf labels.tar in that folder restores labels/ |
donor-studio-resunet-scratch-20260926/{data,config,label-snapshot,qualification,progress}.json |
Frozen run records (also in GitHub) |
donor-studio-resunet-scratch-20260926/checkpoints/resume.pt |
Resumable training state after epoch 9 (model, optimizer, RNG; 74,250 updates) |
donor-studio-resunet-scratch-20260926/hf-epoch-9/ |
Published epoch-9 model.safetensors and inference package (copy of the model repository) |
donor-studio-resunet-scratch-20260926/rebuild/ |
Reports of the 2026-10-10 rebuild (below): select-report.json, plans.tar (per-dataset chunk plans) |
backup/ |
Donor Studio geometry backup (2026-09-26/30): campaign meshes, scan configs, S3 object inventories, renderer recipes and production code, with its own README and checksums; needed only to re-render the CT |
render-runtime/ |
The vc_render_tifxyz build used for all campaign tiles (Docker image ghcr.io/scrollprize/villa/volume-cartographer@sha256:bad516f66001abca759454cc43e4fd11e5b19aa55d36bdc2043817291c8083c4) |
Provenance of the CT tiles (2026-10-10 rebuild)
The run's original prepared/ folders and its local scan copies were deleted from the training host after training.
They were rebuilt with the run's own code:
rebuild_select.pyre-derived every per-dataset plan from the frozen label snapshot with the unchangedprepare.select_one; all 202 eligible datasets match the frozendata.jsonexactly (samples, distinct patches, tiles, shape, supervised ink/background, holdout exclusion); the 9 differing records are datasets excluded from training in the original run too, differing only in free-text exclusion reasons / alias notes.render_rebuild.pyrestored exactly the level-0 chunks each tile reads (the renderer's own chunk plan for campaign tiles) from the public scans — PHerc081420250804134230(9.362 µm), PHerc084120250821151531(9.366 µm), PHerc0009B20250521125136(8.640 µm),s3://vesuvius-challenge-open-data— checking each chunk's size and MD5 against the S3 object inventory recorded in the backup, and then called the unchangedrender.render_row(same renderer binary and image; original-donor rows through the accepted sampler).
Check: the published epoch-9 weights on the rebuilt holdout, scored exactly as train.py's evaluate() (eager instead
of compiled), give loss 0.51900 / AP 0.76566 / ROC-AUC 0.83043 on 1,334 patches (15,791 ink / 29,558 background pixels);
the run logged 0.51891 / 0.76560 / 0.83043 at epoch 9 (rebuild/verify-holdout.json; script verify_holdout.py).
Derived from Vesuvius Challenge open data and released under CC BY-NC 4.0, the licence of that scan data.
- Downloads last month
- 1,185