Prodigle/pherc0306b-ink-ahn

A-hn: ink model for PHerc. 0306B

A-hn is the ink model we trained for our First Letters result on the Herculaneum scroll PHerc. 0306B (Vesuvius Challenge, 2026). The claim image multiplies its ink probability by TRACE v3B's, pixel by pixel. A-hn is the mean of three fine-tunes of v8-in, with seeds 0, 1 and 2.

• Code, method and the claim package: https://github.com/SirProdigle/pherc0306b-first-letters
• Training data, ink maps and the Docker image: https://huggingface.co/datasets/Prodigle/pherc0306b-first-letters
• Liam Warner and Nicola Bodill (papyrologist)

Files

Each checkpoint folder holds model.safetensors, config.json and meta.json. meta.json has the full training config, the data sources and the sha256 of the starting weights. SHA256SUMS covers every file.

folder	model	used in the claim
checkpoints/hn4_s{0,1,2}_final	A-hn, seeds 0, 1 and 2	yes, as the mean of the three maps
checkpoints/hn4_s{0,1,2}_step_0300	A-hn at step 300 of 600	no
checkpoints/hn4_decoy_final	decoy: the same training with every PHerc. 0306B label moved 3.5 × 3.0 mm	control
checkpoints/self_main*	model A: the same training without hard negatives, three seeds	comparison
checkpoints/self_343h_*, hn4B_s0_*	models B and B-hn: A and A-hn with PHerc. 0343 and PHerc. 0139 labels added to the human half	comparison
checkpoints/v8306_343h_s0_*	model C: public human labels only, no PHerc. 0306B data	comparison
checkpoints/v8Q_final	v8Q: v8-in fine-tuned on PHerc. 1447 w058 and the PHerc. 0343 tracings, no PHerc. 0306B data	reading-aid views only
run_records/<run>/	training config, per-step log, tile list and completion record of each run
code/v8in_loader/	the v8-in loader (ink8um) and its config.json

A-hn seed	sha256 of model.safetensors
0	9a2812843924c50cafaa2b598f62c92ceca9f0d1f5d05b7bbf25764814ec7009
1	5c2740c1f33ee5db73ca97425669607a2ffe4ae9af8b9d09d61f1c43165e60ad
2	a294716b083769e8116d5f8d3cfa6a09378ca5c7e51c32225d04703266ad8ce7

Training

setting	value
start	v8-in (YoussefMoNader/ink-8um-v8in, sha256 3b94548d…88c4cd)
data	half of every batch: PHerc. 1447 team labels (w058, w060); half: machine labels on five other PHerc. 0306B surfaces, 26.7 cm², loss weight 0.5
machine labels	mean z-score of v8-in, Reader V3, TRACE v3B and Reader v2; ink where the mean is 2.5 or more in blobs of 0.3 to 6 mm², background where it is 0.3 or less and at least 20 px from ink, the rest ignored; 66.7 mm² of ink in 72 blobs
exclusion	tiles at least 2.9 mm from the claimed line in 3-D; CT inside that buffer set to zero
steps	600 of 16 tiles (micro-batch 4, accumulation 4)
loss	0.5 Dice + 0.5 BCE on labelled pixels, BCE targets smoothed to 0.25 and 0.75
hard negatives	each labelled background pixel's BCE weight is 1 + 4 × the model's own ink probability there
optimiser	AdamW, learning rate 2e-5, 5 % warm-up, cosine decay to 2e-6, weight decay 1e-6
augmentation	the organisers' PHerc. 1447 recipe (rotation, scale, shift, blur, dropout; see meta.json)
compute	about 12 minutes per run on one GPU, peak 19 GB

TRACE v3B, the other half of the claim's product, is also one of the four teachers.

Inference

A-hn reads layers 8 to 31 of a 41-layer render whose layer 20 is the surface, in stored order. It uses 64 px tiles at stride 21 and a ±3σ Gaussian stitch divided by the weight sum, in one pass. The scripts that ran are submission/code/ink_v8.py and submission/code/training/ in the GitHub repo.

Held-out known text

None of these models trained on PHerc. 0841, but TRACE v3B's authors used it to choose their checkpoint, so rows with TRACE v3B are slightly optimistic (about 0.01 AUC by their estimate). False ink is the share of blank papyrus marked as ink at 80 % stroke recall.

model	AUC	AP	false ink
TRACE v3B × A-hn (claim rule)	0.874	0.731	6.1 %
TRACE v3B × v8-in	0.871	0.727	5.9 %
TRACE v3B	0.870	0.713	8.5 %
A-hn	0.845	0.699	12.3 %
v8-in	0.840	0.674	12.1 %

The product of two models gives most of the gain on known text. Records: submission/method/checks/known_text/ in the GitHub repo.

Licence

The weights are CC BY-NC 4.0 and the loader code is MIT (LICENSES/). They derive from the Vesuvius Challenge open data (the PHerc. 0306B and PHerc. 1447 scans and the PHerc. 1447 team labels, all CC BY-NC 4.0) and from v8-in by Youssef Nader (MIT). The teachers were v8-in, Reader V3 and Reader v2 (domenicor046, MIT) and TRACE v3B (Jashann). Models B and C also used the PHerc. 0343 stroke tracings (jeff-j-chen, MIT) and the PHerc. 0139 organiser labels.