jimmylomro/hercunet-corpus

HercUNet training corpus

The pseudo-label training data for HercUNet v0 — everything needed to retrain the model with exactly what we trained on.

• 📦 Code + docs: https://github.com/jimmylomro/hercUNet
• 🧠 Model: https://huggingface.co/jimmylomro/hercunet-v0

Contents

• sample_s*_L*_z*_y*_x*.npz — 4031 stage-1 ∇φ pseudo-label sample bundles (run20260808, 22 scrolls). Each bundle holds the base sheet meshes + the merge/split augmentation records (no dense fields — ∇φ is decoded on the fly at training time). This is exactly the corpus HercUNet v0 trained on.
• corpus_index/corpus_windows.txt — the window coordinate list (SCROLL,Z,Y,X) that regenerates the corpus.
• corpus_index/m7_windows.txt + corpus_index/m7_scout_himat_manifest.json — the mined m7 rehearsal windows (100 coherent m7 surface regions across 20 scrolls) and the exact, deterministic rebuild manifest.

Train with it

# 1) decode the npz sample bundles into an nnU-Net dataset (CT + K candidate crests + owners):
hercunet train export-labels --corpus ./ --out $nnUNet_raw/Dataset301_AffMalisFull --max-candidates 4

# 2) rebuild the m7 rehearsal corpus (deterministic; reads the PUBLIC m7 preds + CT from the open-data bucket):
hercunet labels m7-mine ./m7_corpus --manifest corpus_index/m7_scout_himat_manifest.json --max-candidates 4

# 3) ...then preprocess -> export-owner -> fit  (see docs/training.md in the repo).

You can also regenerate the stage-1 corpus from scratch from the window list (hercunet labels create --old-negatives --coords-file corpus_index/corpus_windows.txt) — see the repo's docs/herculabels.md.

The m7 mined rehearsal corpus — m7_corpus/

m7_corpus/ holds the mined m7 rehearsal cases as ready-to-train nnU-Net tif cases: 100 cases across 20 scrolls (imagesTr/ = CT + 4 sentinel candidate channels, labelsTr/ = {0 bg, 1 surface, 2 ignore}), zlib-compressed. hercunet train export-labels --m7-corpus ./m7_corpus … copies them straight into the training dataset.

These are coherent m7 surface predictions (surface = m7 & normal_coherence > 0.88, incoherent m7 → ignore) mined from compressed regions of the published m7 detector. They are also deterministically reproducible from corpus_index/m7_scout_himat_manifest.json via hercunet labels m7-mine --manifest, which reads the public m7 predictions + CT straight from the Vesuvius open-data bucket (same manifest ⇒ identical labels; the scout has no random seed).

These hosted cases were verified byte-identical (all 100 cases × all channels) to the output of the original research pipeline before publishing.

License: see the GitHub repository.