SharpMeow/pherc0009b-afv-fibers

PHerc0009B papyrus fibers (AFV V2)

Every papyrus fiber that Qual's AFV V2 model finds in the whole PHerc0009B scan, traced into curves with his AVF_SPLINE_GENERATOR: 1,341,032 fibers (837 m of fiber in all), 136,991 of them 1 mm or longer. Horizontal fibers run along the wraps; vertical fibers cross them. Same model, same recipe and the same .afv format as Qual's own published sets, run here on a scan he had not done.

Zoom on dense wraps: horizontal fibers (cyan) follow each sheet, vertical fibers (orange) cross the slice as short marks.	Every fiber in a 256-voxel cube, turning.
Sweeping through 60 slices of the zoom box.	A vertical section through the same place: here the vertical fibers show as long lines.

What is here

File	What	Size
PHerc0009B_afv_v2_1mm.afv	fibers 1 mm and longer, the set most people want: 136,991 fibers, 37,241,628 points	1.22 GB
PHerc0009B_afv_v2.afv	every traced fiber: 1,341,032 fibers, 61,998,860 points	2.64 GB
tiles/	one small .afv per 1024-voxel cell of the model's grid (205 cells with fibers); tiles/index.json gives each cell's level-0 box. Read a box without the whole file.	4.54 GB
stream/	the same fibers as 329 Parquet shards on a spatial grid, every fiber in exactly one, with stream/index.json: stream any box without the whole file (below)	0.78 GB
previews/	the pictures on this card and their views (index.json)
load_example.py	read the fibers in any box with Python's sqlite3 and NumPy
villa-check.jsonl	villa's own AFV reader run over every file here
stats.json	the counts on this card

	Fibers	Fiber length	Median	90th percentile	Longest
All, horizontal	518,770	386 m
All, vertical	822,262	451 m
All	1,341,032	837 m	0.22 mm	1.02 mm	56.7 mm
1 mm and longer, horizontal	69,029	260 m
1 mm and longer, vertical	67,962	251 m
1 mm and longer	136,991	511 m	2.07 mm	8.23 mm	56.7 mm

Get it

hf download SharpMeow/pherc0009b-afv-fibers PHerc0009B_afv_v2_1mm.afv --repo-type dataset --local-dir pherc0009b-afv-fibers
python3 load_example.py pherc0009b-afv-fibers/PHerc0009B_afv_v2_1mm.afv        # or: X Y Z [HALF] in level-0 voxels

Or explore any box, see it over the CT and download just that box as .afv, on the AFV Fibers page.

Stream a box

Read just the fibers in one box, without downloading the whole file, with the Scrolls kit (pip install numpy pyarrow, then from a clone):

python -m kit fibers stream --scan PHerc0009B --box 384 640 4480 4736 4480 4736 --out box.afv

The box is Z0 Z1 Y0 Y1 X0 X1 in level-0 voxels. This one holds 170 fibers and read 2.4 MB of the 2.64 GB whole-scan file (measured 28 Oct 2026). The result is exactly what filtering PHerc0009B_afv_v2.afv gives: the same fibers, and a byte-identical box.afv that villa's importer opens. Add --family H or V, or --min-len 1 for the 1 mm set. Every fiber as rows, streamed with the datasets library:

from datasets import load_dataset
ds = load_dataset("SharpMeow/pherc0009b-afv-fibers", "stream", streaming=True, split="train")
for row in ds.take(3):
    print(row["id"], row["family"], row["length_mm"], len(row["points"]) // 3)

Each row is one fiber: its .afv columns, bounding box (bmin_x to bmax_z), blocks and points (flat x, y, z, float64). Before upload, every test box, a control box outside the scan (no fibers) and the whole scan were streamed back from here and matched the whole-scan file; the format and checks are on the tool page.

Open in VC3D

PYTHONPATH=build/python python3 vc_object_store/scripts/import_afv.py PHerc0009B_afv_v2_1mm.afv pherc0009b-fibers.vcobj \
    --volume-id 20250521125136

Then attach pherc0009b-fibers.vcobj from VC3D's File → Object Store menu, with the 20250521125136-8.640um-1.2m-116keV-masked.zarr volume open (its catalog id must be 20250521125136; see the script's notes for a local cache). villa's import_afv.py takes native level-0 coordinates; every file here is in level-0 voxels of 20250521125136-8.640um-1.2m-116keV-masked.zarr with scale factor 1, and that script's own reader accepted every file before upload (villa-check.jsonl).

How it was made

• Scan: PHerc0009B, 20250521125136-8.640um-1.2m-116keV-masked.zarr, Vesuvius Challenge open data. The model ran on level 0 (8.64 um voxels).
• Fiber maps: Qualzz20/afv_fiber_9um_v2 (nnU-Net, checkpoint sha256 b8519a4c...), Qual's recipe: 128-voxel windows at stride 64 on one global lattice, Gaussian blending, bf16, per-window normalization, no mirroring. 223 blocks of 1024 voxels on Hugging Face Jobs (RTX PRO 6000).
• Curves: AVF_SPLINE_GENERATOR at commit 3c8b7362, run per 1024-voxel region with a 128-voxel halo and seeds only inside the region, on Hugging Face CPU Jobs.
• Stitching: each region's curves cut at its own faces and joined across faces where ends meet within 1.5 voxels (136,337 joins); then the CT exterior cleaning Qual uses (curves touching air connected to the scan's outside are dropped). No region missing.
• Checks: on a test cube of Qual's own published V2 set, this pipeline's fibers match his to 98% within one voxel (recall and precision), with lengths within 1%. Region stitching matches one unsplit run to 99.75 to 99.99% within one voxel. PHerc0009B is not among the 30 scrolls the AFV V2 model card lists as training data.

Notes

Model output, nobody has checked it fiber by fiber. Fibers are geometry, a map of papyrus structure: not a reading of any text.

Related

The AFV Fibers page has all six scans and a catalog of every public fiber set, including Qual's own.

Credit and terms

Fiber model, spline generator and .afv format: Qual, Apache 2.0. Scan: Vesuvius Challenge open data, CC BY-NC 4.0; these fibers are derived from it and shared under the same terms. Runs, stitching and this dataset: Chase Hendrick.