LimeGS/herculaneum-legibility-proxy-labels

Herculaneum legibility proxy — training labels (proxy_v4 / proxy_s2s3_v1)

This is the label set and training recipe behind the proxy_v4 / proxy_s2s3_v1 model card: https://huggingface.co/LimeGS/herculaneum-legibility-proxy. Read that first — it documents the model, the preprocessing, the thresholds, and the validation results. This dataset only covers how the model was trained: the human labeling protocol, the exact windows it produced, and a script to turn them back into pixels and a checkpoint.

proxy_v4 is a ResNet-18 binary classifier that scores ~1 cm windows of Vesuvius Challenge ink-detection maps for legible-text likelihood (connected Greek letterforms vs. fiber noise/damage). proxy_s2s3_v1 is the same model fine-tuned with extra negatives from self-rendered Scroll 2/3 maps.

Files

File	Checksum
train_labels.jsonl	MD5 7af281941a56
fiber_negatives_50.jsonl	MD5 8b9bd89c4a85
full_index_complete.json	SHA256 4d393d70ce886ed62b7e73e365f1d01cbe7f6efa37168fb3f27ade2b89d6e7a8

What's in this release

• full_index_complete.json — the full scoring index this release's model produced over all 78 official Scroll 1 ink maps: 125,298 rows of {panel, y, x, v3}. The score key is named v3 for legacy reasons (the field predates the final checkpoint); every score was produced by proxy_v4.pt. Gold = v3 >= 0.9.
• summarize_index.py — recomputes, with asserts, every published bucket and denominator from that index (125,298 / 15,386 / union 5,678 / 4,424 = 77.9% / classics 2,460 / redundant 7,248).
• eval_checkpoint.py — frozen evaluator: deterministically reconstructs the held-out validation split (81 positives / 123 negatives) and scores the published checkpoint. Expected: AUROC 0.9853457794.
• fiber_negatives_50.jsonl — the 50 GPU-verified-fiber negative windows (weight 1.0) the original proxy_v4 run trained on: the top-50, by score, of the project's full-grid candidate ranking. Same row schema as the main file; 12 of the 50 fall in the validation panels and train.py routes them there, as in the original run.
• train_labels.jsonl — 686 rows, one JSON object per line. Every row is a {panel/segment, y, x, window size} coordinate into an already-public CC BY-NC 4.0 map, plus the human (or human-implied) verdict, a training weight, and full provenance back to the original labeling-session file.
• generate_crops.py — turns those coordinates back into the exact 160×160 training crops, using the identical preprocessing documented in the model card (percentile [2, 99.5] stretch over active pixels → resize → pad for undersized fragments).
• train.py — the consolidated training loop (ResNet-18, BCE-with-logits, AdamW, per-example weighting, spatial holdout) that turns crops back into a checkpoint with the same architecture as the published one.
• requirements.txt — the pinned dependency versions this release was built and smoke-tested with.
• test_package.py — self-checks (row counts, schema, id uniqueness, the 12-in-validation fiber invariant, and the index SHA256); runs from any directory with no network.
• VERIFICATION.md — this release's own QC pass: every row count checked against the source files, and checked against the model card's prose claims (with one confirmed discrepancy — see below).

Why no images are bundled

The label rows are coordinates, not pixels. Every map they point into is either:

1. an official Vesuvius Challenge ink-detection map (Scroll 1 / PHercParis4, the June-2026 w010–w100 panel series), CC BY-NC 4.0, downloadable from the public anonymous S3 mirror; or
2. a self-rendered Scroll 2/3 mosaic produced by this project's own rendering pipeline (used only because no official ink map existed yet for those regions) — see "Known limitation" below for what that means for reproducibility.

Publishing coordinates instead of crops keeps this release at a few hundred KB instead of redistributing gigabytes of imagery, while still making the exact training set fully specified and (for the S1 portion) fully regenerable by anyone with access to the public maps.

Labeling protocol

Scroll 1 (PHercParis4), three rounds, all against the 16 official ds8 w010–w100 panels, 512×512 px windows (≈1 cm physical):

1. browse_all (human_marks_v1.json, 398 rows) — a reviewer paged through model-agnostic candidate windows and tagged each interesting or unsure. 393 interesting → positive; 5 unsure → excluded.
2. model_proposed (round2_top120.json / round2_confirmed.json, 120 rows) — the top-120 windows by an early classifier's score were shown to the reviewer for confirmation. All 120 were confirmed (0 rejected) → positive. (round2_top120.json is the pre-confirmation candidate list; round2_confirmed.json — the file the training scripts actually read — has the identical 120 coordinates plus a human: "confirmed" field, so both are represented by one row each, sourced from round2_confirmed.json.)
3. uncertainty_sampled (round3_marks.json / round3_uncertain.json) — 132 windows in the model's ambiguous score band (0.35–0.70-ish) were pulled as round3_uncertain.json; the reviewer explicitly tagged 45 of them in round3_marks.json (28 interesting → positive, 17 unsure → excluded). The other 132 − 45 = 87 were shown to the reviewer during the same pass but never flagged — treated as implied negatives at weight 0.7 (round uncertainty_sampled_implied, source round3_uncertain.json), on the reasoning that a genuinely interesting window in a browsed set would have been flagged.

Additionally, verification_only (verify6.json, 6 rows) — six already-gold (score 1.0) windows that were independently cross-checked against a native-resolution CT render on a different pipeline (tools/session-20260706/verify6_driver.sh) as a sanity check, not a labeling round. This file has no interesting/negative verdict field and is never read by either training script — included here for transparency of everything that touched these maps, marked split: excluded, label: unsure. Two of the six coincide exactly with an existing round-1/2/3 positive coordinate (flagged per-row in the note field).

Scroll 2/3, one round (context_reviewed, s2s3_context_review_final.json, 30 rows): a first pass (s2s3_seed_round1_marks.json) tagged 30 candidate windows from the self-rendered S2/S3 mosaics as 7 interesting / 23 unsure. Because these were sparse, low-signal fragments, every one was re-reviewed with 3×3 spatial context added around it; most downgraded to noise. Final: 23 noise (→ negative, weight 1.0), 3 interesting + 4 unsure (→ excluded, never trained on — these are the only S2/S3 windows that stayed live candidates after context review, and none were confirmed as actual text).

Row counts (verified against the actual files — see VERIFICATION.md for the full derivation)

Scroll	Round	Rows	Label	Split	Weight
S1	browse_all	393	positive	train	1.0
S1	browse_all	5	unsure	excluded	—
S1	model_proposed	120	positive	train	1.0
S1	uncertainty_sampled	28	positive	train	1.0
S1	uncertainty_sampled	17	unsure	excluded	—
S1	uncertainty_sampled_implied	87	negative	train	0.7
S1	verification_only	6	unsure	excluded	—
S1 total		656	541 positive / 87 negative / 28 excluded
S2/S3	context_reviewed	23	negative	train	1.0
S2/S3	context_reviewed	7	unsure	excluded	—
S2/S3 total		30	0 positive / 23 negative / 7 excluded
Grand total		686

Known limitation — this is not the complete negative set the published checkpoints trained on

The label rows above cover every window that a human explicitly looked at. The actual training runs (train_proxy_v4.py, train_proxy_s2s3_v1.py in the project's internal archive) also mixed in two more negative sources that are not human labels and so are deliberately not in train_labels.jsonl:

1. 50 GPU-verified-fiber windows (weight 1.0) — the top-50 of an automatically-scored full-grid candidate ranking over all S1 windows (a model/heuristic output, not a human judgment — which is why they ship separately from the human labels). Included in this release as fiber_negatives_50.jsonl (2026-07-11 peer-review fix: an earlier version omitted them, which made the published AUROC non-reproducible since 12 of the 50 sit in the validation panels). Provenance caveat: the original script read a session copy of the ranking; the shipped file applies the same selection rule (first 50 of the score-sorted ranking) to the archived ranking file.
2. Procedural background negatives (~30 per panel that contains a positive mark, weight 1.0, fixed seed) — a content-thresholded random sample with no human judgment involved at all, fully regenerable from the raw map pixels alone. train.py regenerates this automatically from --maps-dir (same directory used for crop generation), reproducing the original scripts' algorithm exactly.

Separately, and more importantly for proxy_s2s3_v1: the model card states "192 additional negatives (23 human-reviewed noise at weight 1.0, 169 implicit at 0.7)". This release accounts for the 23. The 169 implicit negatives' coordinates could not be recovered — train_proxy_s2s3_v1.py reads them from a ~199-item candidate gallery (gallery_items, loaded via an S2S3_ITEMS environment variable pointing at a session-specific scratchpad path) that no longer exists anywhere in the project repo or filesystem. Only the 30 windows that a human actually reviewed (s2s3_context_review_final.json) survived. This means:

• train_labels.jsonl is a complete, verified record of every S2/S3 window a human looked at.
• It is not sufficient to exactly reproduce proxy_s2s3_v1.pt's training run — train.py --mode s2s3 will fine-tune on 23 S2/S3 negatives instead of 192, and should be expected to land close to but not identical to the published checkpoint's behavior on self-rendered maps. See VERIFICATION.md for the full accounting.

How to reproduce the model end to end

# 0. Verify the PUBLISHED checkpoint first (deterministic; expected
#    AUROC 0.9853457794 on 81 positives / 123 negatives):
python generate_crops.py --labels train_labels.jsonl \
    --maps-dir /path/to/s1_official_panels --out ./crops
python generate_crops.py --labels fiber_negatives_50.jsonl \
    --maps-dir /path/to/s1_official_panels --out ./crops
python eval_checkpoint.py --checkpoint /path/to/proxy_v4.pt \
    --maps-dir /path/to/s1_official_panels --crops-dir ./crops

# 1. (optional, for the s2s3 variant) S2/S3 crops need the self-rendered
#    harvest (project-internal, not redistributed — see "Known limitation"):
python generate_crops.py --labels train_labels.jsonl \
    --maps-dir /path/to/s1_official_panels \
    --s2s3-dir /path/to/s2s3_harvest --out ./crops

# 2. Retrain a proxy_v4 equivalent (ImageNet init; labels + fiber negatives
#    + procedural background negatives, original spatial holdout):
python train.py --mode s1 --labels train_labels.jsonl --crops-dir ./crops \
    --maps-dir /path/to/s1_official_panels \
    --extra-negatives fiber_negatives_50.jsonl --out proxy_v4_reproduced.pt

# 3. Fine-tune the proxy_s2s3_v1 equivalent (see "Known limitation" above —
#    won't be bit-identical to the published checkpoint):
python train.py --mode s2s3 --labels train_labels.jsonl --crops-dir ./crops \
    --maps-dir /path/to/s1_official_panels \
    --init-checkpoint proxy_v4_reproduced.pt --out proxy_s2s3_v1_reproduced.pt

# 4. Compare: outputs are plain resnet18()+Linear(512,1) state_dicts,
#    122 tensors, same keys/shapes as the published checkpoints. Score your
#    retrained model with eval_checkpoint.py --checkpoint <yours>; expect
#    close to, not exactly, the published 0.9853457794.

Trains in minutes on a laptop (Apple MPS, CUDA, or CPU) — this was always meant to be a small, cheaply-reproducible model, not a reason to withhold the labels.

Provenance & license

Labels are the project's own human review, over coordinates into publicly released Vesuvius Challenge data (CC BY-NC 4.0). Released under the same CC BY-NC 4.0 to match the model card. No official-team code, checkpoints, or imagery are redistributed here — only coordinates and a from-scratch training script.