nestorvfx/vesuvius-ink-training

Vesuvius ink — training-ready

Inputs and splits for a general ink-detection model that works on any scroll: a 3-D encoder pretrained on the scrolls' own CT, fine-tuned on the official labels. Every input here was made by the same code a new scroll goes through — its CT sampled along its traced surface at native resolution, straight from the official open-data volumes — and checked against the official render of the same surface (per-tile correlation in each stack.json).

Labelled area in this release: 130.5 cm² supervised, 34.1 cm² ink.

Contents

Part	Folders	Objects	Pitches (µm)	Size
surface stacks + labels	33	7	2.215, 2.399, 2.4, 2.401, 2.403, 9.362	296.3 GB
Total	33			296.3 GB

Layout

stacks/<sample>/<segment>/<pitch>um/stack.zarr        [depth, H, W] uint8, 1-voxel steps along the
                                                      surface normal, centred on the surface
stacks/<sample>/<segment>/<pitch>um/labels/*.zarr     ink labels and masks on the same canvas
stacks/<sample>/<segment>/<pitch>um/geometry/         the official atlas mesh (tifxyz) the stack follows
stacks/<sample>/<segment>/<pitch>um/tiles.json        tile index (ink / supervised fraction per tile)
stacks/<sample>/<segment>/<pitch>um/stack.json        provenance: mesh, volume, crop, check vs official
pretraining/                                          exact chunk lists (key, size, ETag) of the 36
                                                      unlabelled survey volumes and of the voxels
                                                      around every labelled surface — fetched from S3
                                                      when (continued) pretraining, never stored here
splits/SPLITS.json, splits/SPLITS_REPORT.md           whole-object splits and their leakage checks
records/                                              physical-surface table, duplicate-surface scan,
                                                      validation-mask audit, the build's state record
manifest.jsonl                                        one row per stack

Splits

The unit of every split is the object (a scroll with all its fragments and pieces), never a tile or a region: leave-one-object-out, the IR-verified fragments as the test set, and a production split. Held-out objects are excluded from pretraining too. See SPLITS_REPORT.md.

Licence and citation

CC BY-NC 4.0 (derived from the Vesuvius Challenge open data). See LICENSE.md. Cite:

Giorgio Angelotti, Stephen Parsons, Sean Johnson, Elian Rafael Dal Prà, Johannes Rudolph, Paul Tafforeau, Alessandro Mirone, Paul Henderson, Hendrik Schilling, Forrest McDonald, David Josey, Youssef Nader, C. Seth Parker, W. Brent Seales. Vesuvius Challenge - CT Scans of Herculaneum Papyri. Vesuvius Challenge.

Processed with ink_corpus (the processing code; every step is recorded in each folder's JSON records). Card generated 2026-09-24.