Prodigle/pherc0306b-first-letters

PHerc. 0306B First Letters: training data, ink maps and Docker image

The data behind our First Letters result on the Herculaneum scroll PHerc. 0306B (Vesuvius Challenge, 2026; scan volume 20250521133212, 8.64 µm, 116 keV). It has every stage of the ink model's training data, the claim's ink maps and render, and the Docker image that reproduces the claim image.

• Code, method and the claim package: https://github.com/SirProdigle/pherc0306b-first-letters
• Model weights: https://huggingface.co/Prodigle/pherc0306b-ink-ahn
• Liam Warner and Nicola Bodill (papyrologist)

Training data, by stage (data/)

folder	what
0_exclusion_zone/	The zone: 2.39 million 3-D points of every surface we have shown of the claimed line. It comes with its build and check code (build_zone.py, check.py), the list of all 23,331 training tiles, and their audit (exclusion_audit.json; all allowed, the nearest 2.925 mm away).
0_inputs/	The five other PHerc. 0306B surfaces the machine labels were made on: tifxyz meshes and 41-layer CT renders (regions_pool.json). The renders can be made again from the meshes and the public volume.
1_pool_masks/	Per surface: the distance to the zone, the on-sheet score and the CT coverage.
2_teacher_maps/	The maps of the four public models on those surfaces: v8-in, Reader V3, TRACE v3B and Reader v2. CT inside the zone's buffer was set to zero first.
3_machine_labels/	The machine labels. Training used main; sc (sharp core) is a stricter variant, and decoy holds the shifted labels of the decoy control. Also here: statistics (stats.json) and overview JPGs.
3b_cloud_packs/	The exact arrays the seeds trained from (306_*.npz), plus the PHerc. 1447 w058 and w060 anchor packs, re-hosted from YoussefMoNader/ink-8um-pherc1447-surfaces (CC BY-NC 4.0).

The label rule:

1. Average the four teachers' z-scores.
2. Mark ink where the mean is 2.5 or more, in blobs of 0.3 to 6 mm².
3. Mark background where the mean is 0.3 or less, and at least 20 px from ink.
4. Ignore everything else.

Results (results/)

file	what
claim_surface_render_41layers.zarr	The 41-layer CT render of the claim surface PHerc0306B_20250521133212_w074own_touchup_V5Rc. Layer 20 is the surface, and layers are 1 voxel apart.
ct_surface_layer20.png	Its surface layer, as used under the ink in ink_over_CT.jpg.
member_maps/	The per-model ink maps on that render: TRACE v3B (fwd_trace_v3B.npy), the three A-hn seeds (fwd_hn4_s*.npy) and the decoy; Reader v2, Reader V3, v8-in (±1σ and ±3σ stitch) and v8Q for the reading-aid views. The shuf_* files are the same maps on the depth-shuffled render.
fused/	The claim rule's outputs: fwd_prod.npy is TRACE v3B × A-hn, and fwd_ensA.npy is the A-hn mean. fuse.json records every input's sha256.
claim_map_fwd_fin.npy, depth_shuffle_map_fwd_fin.npy	The claim map and its depth-shuffle control.
known_text/	Held-out known-text scores of the claim rule.

All maps are on the render raster (4900 × 7780 px). The claim frame is u = −36 + x / 115.741 mm and v = −21 + y / 115.741 mm.

Docker image (docker/)

fl-claim-a7_final.tar.zst (6.62 GB, sha256 1832832330ab67b134795eaedd43d6d5c3dcf004596e1fd287d2e71b3af73536) loads with zstd -dc fl-claim-a7_final.tar.zst | docker load. It redraws the claim image byte for byte, re-runs the three A-hn seeds on part of the CT, and re-checks every training tile against the exclusion zone. docker/README.md has the commands, and docker/THIRD_PARTY_NOTICES.md lists the licence of everything inside, including the GPLv3 VC3D AppImage. docker/build/ holds the files that only a from-scratch build needs (docker/build.sh in the GitHub repo): the VC3D AppImage, whose release is no longer on GitHub, the smoke-test CT crop, two files of the training tree and the reference views.

Licence

The data is CC BY-NC 4.0 and the code in 0_exclusion_zone/ is MIT (LICENSES/). It derives from the Vesuvius Challenge open data (the PHerc. 0306B scan and the PHerc. 1447 scan, renders and team labels, all CC BY-NC 4.0). The teachers were v8-in (Youssef Nader, MIT), Reader V3 and Reader v2 (domenicor046, MIT) and TRACE v3B (Jashann). SHA256SUMS covers every file.