SharpMeow/first-letters-ink-atlas

First Letters ink atlas

Seven public flat ink models, run on the same public traced surfaces of the First Letters scrolls, with every map kept: both depth orders, three depth windows, and a z-score ensemble. Each surface is rendered once on one shared 9.6 µm canvas, so maps from different models line up pixel for pixel and can be compared, stacked or ensembled directly.

This release has 32 surfaces: 25 surfaces from 13 First Letters scrolls and 7 reference surfaces on PHerc0343. Every map is model output, not a reading.

What is in it

Per surface
maps/MODEL/WINDOW_ORDER.png	each model's map, 8-bit (0 to 1 as 0 to 255), windows c-4, c+0, c+4 (center layer, in 9.6 µm layers; Hecate 9.6 has the center window only), orders forward and reverse
ensemble/ORDER.png	z-score mean of the alive models, the stored order read as the reading side, z from -2 to 4 as 0 to 255
ensemble-rev/ORDER.png	the same with the reverse order as the reading side
ct_mid.png	the middle CT layer of the render
preview.jpg	CT and ensemble side by side
meta.json	mesh and scan links, render record with its SHA-256, each model's liveness at every window and order, training-exposure note, the face check of every model pair

atlas.json lists every surface with its path, and the surface counts per scroll.

How it was made

1. Surfaces. Traced meshes from ketchupbyheinz's vesuvius-scroll-survey (commit 1f12c1ce), plus the published segments in the open-data bucket.
2. Render. Each mesh is rendered from level 0 of the public CT, streamed chunk by chunk, onto a 9.6 µm canvas with 65 layers 9.6 µm apart (trilinear sampling, the same canvas rule vc_render_tifxyz uses with --scale). One canvas for every scan, whether 8.64 or 9.362 µm, so each 9 µm model reads its own scale.
3. Models. All seven run through the Ink Lens comparison engine, with pinned weights and SHA-256 checks, villa's own flat inference and ink-sdk's Hecate adapter. A liveness check marks a model alive on a surface when its map has real spread; 25 surfaces have all seven alive and 7 have six.
4. Ensemble. ink-sdk's ensemble, Erwin Nieuwlaar's PHerc0343 recipe: the mean of per-model z-scores.
5. Face check. For every pair of alive models, ketchupbyheinz's face check from villa PR #2318 scores whether their strongest marks agree better on one face than on the other.
6. Release rule. A rule fixed before any map existed decides which surfaces go out in this release. It had to flag the PHerc0343 mesh behind the First Letters prize awarded on 8 October 2026, and it did; it flagged all seven PHerc0343 references. Surfaces not in this release may be added later.

Models

Model	By	License
ink_9um seeds 42 and 43, step 75,000	Vesuvius Challenge	MIT
Reader v2	Domenico Russo	MIT
Reader V3	Domenico Russo	MIT
dense_native, step 16,000	Erwin Nieuwlaar	MIT
d9v2	TAUIL-Abd-Elilah	MIT
Hecate 9.6 µm	Vesuvius Challenge	MIT

Training exposure: none of these models' published training lists names any scroll in this release, PHerc0343 included (d9v2, Reader v2 and Reader V3 list PHerc0343P, a separate object). Hecate's audited training manifest is not published, so its list rests on its committed trainer, and one fragment id of its base model is unidentified; each meta.json records the exposure of every model as the published sources state it.

Use it

from huggingface_hub import snapshot_download

root = snapshot_download("SharpMeow/first-letters-ink-atlas", repo_type="dataset",
                         allow_patterns=["atlas.json", "surfaces/PHerc0268/*"])

Each meta.json links the mesh (mesh.base) and the CT (ct), so a surface can be opened in VC3D or neuroglancer from the same files the atlas used.

Credits and license

• CT scans: Vesuvius Challenge open data, CC BY-NC 4.0.
• Traced meshes and the face check: ketchupbyheinz (vesuvius-scroll-survey, code MIT, meshes CC BY-NC 4.0; villa PR #2318, MIT).
• Models: as in the table above. Ensemble recipe: Erwin Nieuwlaar, through LimeGS's ink-sdk.
• Atlas: Chase Hendrick, with Ink Lens.

The maps derive from CC BY-NC 4.0 scans and meshes, so this dataset is CC BY-NC 4.0.