domenicor046/reader-v2

Reader v2: beats every public 9 µm ink model on the 116 keV First Letters scan protocol

A drop-in for the team's own ink_9um inference command: only the checkpoint changes. Held out: AUC 0.834 vs 0.725 for the next best model on 116 keV scans (the protocol of 11 of the 22 First Letters scrolls), 5× less false ink than Hecate at the same ink recall (10.4% of blank papyrus vs 58%), and on PHerc0841, a scroll it never saw, the highest AUC of any public model against the team's map (0.858). Code, figures and every evaluation: github.com/DomRusso2/reader-v2.

hf download domenicor046/reader-v2 reader-v2-step040000.pth --local-dir .
python -m koine_machines.inference.infer <surface_volume.zarr> reader-v2-step040000.pth out.tif \
  --overlap 0.5 --blend-mode hann --direction forward

(koine_machines = ink-detection/ on villa's merge-ink-pipelines branch; add --no-compile on Windows.)

file	what it is
reader-v2-step040000.pth	the model (model / config / step, 138 MB, same format as scrollprize/ink_9um). sha256 654ec5acec2b6c4788d9cb326e2f0b8c2c730584949a934ef775db7f6ab5d3a6
reader-v2-init-ft12k.pth	the training init (my August ink9um-dense checkpoint fine-tuned 12k steps at 116 keV), for retraining with train_config.json. sha256 6bad92971b029857f23d1ab46a623b0f2cd6c9605c05d1939c8f859b0d98f79b
train_config.json	the exact training config (paths relative to the project root)

Scoreboard (held out; same data, same masks, same metric for every model)

held-out test	Reader v2	released ink_9um (best of 14)	best other public model
116 keV PHerc0009B, AUC vs team map (4 segments)	0.834	0.626	0.725 (Hecate)
116 keV PHerc0009B, AUC vs human labels	0.927	0.813	0.901 (Nieuwlaar)
False alarms on blank papyrus (at 80% ink recall)	10.4%	43%	10.5% (Nieuwlaar)
Never-seen scroll PHerc0841, AUC vs team map	0.858	0.708	0.848 (Hecate)
113 keV PHerc0139 w047, AUC	0.893	0.778	0.898 (Nieuwlaar)
Never-seen scroll PHerc0841, AUC vs human labels	0.824	0.733	0.855 (Hecate)

First on four of the six tests, second on the other two, and ahead of the team's released checkpoint on all six.

Limits: Hecate alone leads on PHerc0841's human labels (averaging the two gives 0.866; script in the GitHub repo); one training seed; teacher maps are model outputs from finer scans; on PHerc1447 it shows one letter-like form, not readable lines.