LimeGS/herculaneum-legibility-proxy-labels-v5

Herculaneum legibility V5 — training data and recipe (R2B, EXPERIMENTAL)

This is the cohort, splits, source bundles, raw crops and full code behind the V5 scorer: https://huggingface.co/LimeGS/herculaneum-legibility-proxy-v5.

V5 is a new recipe that improves on the published proxy_v4 (https://huggingface.co/LimeGS/herculaneum-legibility-proxy). It retrains a ResNet-18 on a corrected historical cohort and scores macro AUROC 0.810 vs 0.756 for proxy_v4 (leave-one-scroll-out, four scrolls, three seeds). The exported checkpoint is the predeclared candidate R2B, seed 0, marked EXPERIMENTAL.

Read evidence/REPORT.md and evidence/results_independent_review.md before using anything in this repo. (Both are written in Spanish; this card and the code comments are in English.)

Files

Path	What it is	SHA-256 (first 12)
data/samples.jsonl	2,836 rows in tensor order: index, sample_id, physical_scroll, source, y, image_sha256, pixel_sha256, input_tensor_sha256, training_only, label_status, physical_group	1bd9ef6352b9
data/labels.npy	int64 [2836], same order as samples	dcd035c138ce
data/features.npy	float64 [2836,4] original normalized mean/std/p90/p10 (OLS features, not CNN-stretched)	dbf1af5f5fe0
data/splits.json	5 leave-one-scroll-out folds: train_r1, train_r2, validation, test, calibration, plus test_scroll, validation_scroll and u (tensor-row indices)	4e7f31f46c19
data/dataset.json	schema, n=2836, recipe stretch160, hashes of the four files above	186825041dbd
sources/inputs.jsonl	2,406 baseline rows (image path, hashes, labels, campaign metadata)	d8708d0aa322
sources/baseline_tensors.npy	uint8 [2406,160,160] canonical baseline tensors	05eacd3554a8
sources/baseline_samples.jsonl	2,406 baseline prepared rows	2ca0c044229f
sources/all_binary-bundle-v2.json	full source bundle: 2,406 baseline + 430 additions with evidence	385399db4738
crops/	the 2,406 baseline + 430 additional raw crops (structure preserved)	—
code/	prepare_data.py, preprocess.py, train.py, evaluate_training.py, export_model.py, score.py, training_config.json, refit_config.json	—
evidence/	REPORT.md, results_independent_review.md, freeze.v2.json, results/comparisons.json, results/results_per_roll_metrics.png, export_cpu_smoke.json	—
test_package.py	offline self-checks (row counts, hashes, index order, no path leaks)	—

The cohort: 2,198 positives / 638 negatives, 2,406 corrected historical originals + 430 training-only additions (404 provisional majorities + 26 historical consensus). All physical_group are null (unknown). No new independent human certification is claimed.

Label semantics. y is the binary target the classifier is trained on: y=1 = a window the project rated as containing legible letterforms, y=0 = fiber texture / noise / damage. It is a legibility label, not an ink-presence label. label_status records provenance:

label_status	Meaning
corrected_historical_binary	a historical label after the September-2026 campaign-policy remediation (per-vote cutoff, explicit tie abstention) — the 2,406 originals
provisional_majority	a majority vote over historical reviewers, not independently certified (404 additions)
consensus_final	historical reviewer consensus, still training-only (26 additions)

source is baseline (originals) or additional (the 430 training-only additions); training_only is true for the additions, which never enter validation or test.

Reproduction

data/samples.jsonl + data/labels.npy are not enough to rebuild the pixels: prepare_data.py regenerates tensors.npy (uint8 [2836,160,160]) from the four sources/ files and the raw crops in crops/, and it verifies every file's SHA-256 before proceeding. The structure expected is:

• baseline crops: crops/<sub>/... matching the image_path in sources/inputs.jsonl;
• additional crops: crops/additional/<file> matching derived_crop_path in sources/all_binary-bundle-v2.json.

Note: sources/ files carry local absolute workstation paths (the original checkouts the pipeline ran from) as provenance evidence; they are not portable as-is. data/ (the shipped cohort) carries no local paths or reviewer identities. code/training_config.json and code/refit_config.json likewise carry local reference paths for the historical F0 cache.

sources/baseline_tensors.npy is included so the baseline canonical tensors do not need to be re-derived from pixels; prepare_data.py checks the regenerated baseline against it row by row. The regenerated tensors.npy (72 MB) is not included — it is deterministic from the above.

Running prepare_data.py outside the original workstation. The image_path / derived_crop_path values inside sources/inputs.jsonl and sources/all_binary-bundle-v2.json are absolute original-workstation paths, and those JSONL files are hash-pinned, so editing them breaks the source checks. prepare_data.py resolves each crop path as-is. Two prefixes appear:

• …/legibility-audit-2026-09-04/data/crops/ (2,406 baseline rows) — the shipped crops/<sub>/... tree has exactly the content this prefix expects.
• …/legibility-evidence-recovery-2026-09-04/labels/additional_crops/ (430 additions) — the shipped crops/additional/ holds those files.

To reproduce, recreate those two directories (or symlink them) pointing at the shipped crops/ and crops/additional/ content, then run the command below. The --inputs/--baseline-* / --source-bundle flags only point at the pinned files themselves; they do not rewrite the crop paths inside them. Every crop is SHA-256-checked, so any mismatch fails loudly rather than silently scoring wrong pixels.

python3 code/prepare_data.py \
  --inputs sources/inputs.jsonl \
  --baseline-tensors sources/baseline_tensors.npy \
  --baseline-samples sources/baseline_samples.jsonl \
  --source-bundle sources/all_binary-bundle-v2.json \
  --out /tmp/v5_repro

Training, evaluation and export:

python3 code/train.py run --data <data dir> --config code/training_config.json \
  --init <resnet18-f37072fd.pth> --out <runs dir> --device cuda \
  --expected-gpu-uuid <uuid>
python3 code/evaluate_training.py --data <data dir> --config code/training_config.json \
  --runs <runs dir> --out <results dir>
python3 code/export_model.py --refit-dir <refit dir> --refit-config code/refit_config.json \
  --evaluation <results>/comparisons.json --freeze evidence/freeze.v2.json --out <export dir>

code/train.py needs the pinned ImageNet V1 initializer (resnet18-f37072fd.pth, sha256 f37072fd47e89c5e827621c5baffa7500819f7896bbacec160b1a16c560e07ec) and an authorized CUDA device; the runner enforces the original research GPU UUID. This is not a generic-server recipe.

Verification

python3 test_package.py   # offline, no network

Checks the four data/ file hashes, n=2836, index order, label/y agreement, 2,198/638 split, null physical_group, and that samples.jsonl carries no local absolute paths.

Verified locally 2026-09-29: code/prepare_data.py reproduces samples.jsonl, labels.npy, features.npy, splits.json and dataset.json bit-for-bit from sources/ + crops/, and regenerates tensors.npy matching its declared hash (a817b18c4cbc67586346c57eaa58b22aa2289e233d800aeb8620607912bc5aef).

Limits

Scope is historical_unharmonized_diagnostic. Missing: physically transported masks, certified crop/raster/geometry correspondence, fine territorial groups, and a new independent confirmatory test. All physical_group are unknown, so there is no territorial confidence interval. The 430 additions are not a newly certified sample. The final all-data refit has no held-out evaluation of its own. This is not independent validation, and scores are uncalibrated ranking scores, not probabilities; the artifact is labelled EXPERIMENTAL for that reason. See evidence/REPORT.md §"Qué queda por demostrar" (Spanish) and evidence/results_independent_review.md.

License

Data released under CC BY-NC 4.0, matching the Vesuvius Challenge ink maps they derive from. The label rows contain no reviewer identities and no local paths.