README.md

metadata

license: cc-by-nc-4.0
tags:
  - vesuvius-challenge
  - herculaneum
  - ct
  - lasagna
  - lasagnetta
  - ome-zarr
pretty_name: PHercMANB Lasagnetta field (a preview of villa's lasagna field)

PHercMANB Lasagnetta field

A Lasagnetta preview of villa's lasagna field for the whole PHercMANB scan, in villa's own layout, so you can point lasagna or the spiral fit at it without running a model yourself. The open data has no lasagna field for this scroll.

Get it

hf download SharpMeow/phercmanb-lasagnetta --repo-type dataset --local-dir phercmanb-lasagnetta
cd phercmanb-lasagnetta && sha256sum -c SHA256SUMS
for f in parts/*.tar; do tar -xf "$f"; done
tar -xf PHercMANB-lasagnetta-meta.tar   # last: the whole-scan PHercMANB.lasagna.json and preview.json

You end up with PHercMANB.lasagna.json and PHercMANB_{cos,grad_mag,nx,ny}.ome.zarr, 33.0 GB in all.

Load it

import json
import numpy as np
import zarr

meta = json.load(open("PHercMANB.lasagna.json"))
cos = zarr.open(meta["groups"]["cos"]["zarr"], mode="r")     # cos, 2x down from the CT level read
nx = zarr.open(meta["groups"]["nx"]["zarr"], mode="r")       # normals, 4x down; stored as v * 127 + 128
ny = zarr.open(meta["groups"]["ny"]["zarr"], mode="r")
z, y, x = 245, 1018, 878                                      # a voxel on the normals' grid that holds data
a, b = (float(nx[z, y, x]) - 128) / 127, (float(ny[z, y, x]) - 128) / 127
c = np.sqrt(max(0.0, 1 - a * a - b * b))                     # nz is not stored
patch = cos[2 * z, 2 * y:2 * y + 64, 2 * x:2 * x + 64]       # the same place on the cos grid
print("normal", round(a, 3), round(b, 3), round(c, 3), "cos patch mean", float(np.mean(patch)))

load_example.py is this snippet; it was run on an unpacked part before upload.

What it is

• Scan: PHercMANB, 20260323091048-2.399um-0.2m-78keV-masked.zarr, read at level 2 (2.399 um scan, 9.6 um voxels at that level), Vesuvius Challenge open data
• Model: Lasagnetta 0.1.1, weights at revision 0f9ad8cd: the release recipe (three students averaged, cos blur 0.5, cos offset 0.04)
• Code: ChaseHendrick/Scrolls scripts/lasagnetta at a011d5e4, tile 320 x 576 x 576
• Channels: cos at OME-Zarr level 3 to 4 (2x down from the level read); grad_mag, nx, ny at level 4 (4x down)
• Run: 26 parts on Hugging Face Jobs (RTX PRO 6000), one x, y, z box each. Every part edge sits on a multiple of 512 scan voxels, so every chunk of every level was written by exactly one part, and the parts together are the whole field. Values are kept only where the masked scan holds CT; everywhere else is fill 0.
• Previews: xy at z about 37,664 and xz at y about 13,360 (full-resolution voxels), from OME-Zarr level 4.

How close to villa

Lasagnetta copies villa's lasagna net and runs about 5 times faster. On 30 crops of this scan that no student had trained on, it matched villa's sheet crossings at peak F1 0.870 (crossings matched within one cos voxel; 25 Oct 2026). The model card has the method and the pooled numbers. The release's self-check, run on this whole volume, kept Lasagnetta here (median peak F1 over its spots 0.82 or more).

Notes

Model output: a preview of another model's output, not a reading of any text, and nobody has checked it by eye. villa's own field stays the reference where exact values matter.

The scan also holds a second, less dense object (the upper one in the xy preview). There the scan shows few layers, cos is nearly flat, and it changes level at some part edges.

Related

The Lasagnetta guide shows how to run it on your own region, with a self-check that tells you where to use villa's net instead. villa lasagna fields for ten more scans are on the Lasagna Fields page.

Credit and terms

The scan is Vesuvius Challenge open data, CC BY-NC 4.0; this derived field is shared under the same terms. Lasagnetta copies villa's lasagna (ScrollPrize/villa, weights scrollprize/lasagna, MIT); credit to its authors. Shared by Chase Hendrick.