domenicor046/reader-v2-labels

Dense native-resolution ink labels for 9 µm training (Reader v2)

The training labels of Reader v2: 70 segments of six objects, on the native 8–9 µm scans (9.362 µm / 113 keV, 8.64 µm / 116 keV, 7.91 µm / 53 keV), made from the team's own ink maps of a finer scan of the same segment. Same zarr layout and parameters as the team's ink_9um labels, so a training config picks them up by changing one path.

object	native scan	segments	teacher map
PHerc0139	9.362 µm, 113 keV	20	2.399 µm, 78 keV
PHerc0814	9.362 µm, 113 keV	8	2.399 µm
PHerc0500P2	9.362 µm, 113 keV	18	2.215 µm
PHerc0009B	8.64 µm, 116 keV	14	2.401 µm
PHerc0343P	8.64 µm, 116 keV	6	2.215 µm
PHerc0172	7.91 µm, 53 keV	4	same-scan team map

hf download domenicor046/reader-v2-labels --repo-type dataset --local-dir .
tar -xf reader-v2-dense-native-labels.tar

That gives reader-v2-labels/<object>/<segment>/<segment>_inklabels.zarr and <segment>_supervision_mask.zarr (zarr v2, uint8, shape = the native surface volume (depth, H, W), chunks (depth, 128, 128), 255 = ink / supervised, labels on the middle layer depth // 2). manifest.csv gives, for every segment, the open-data key of the native surface volume the labels sit on, the team map they came from, and the labelled fractions. The surface volumes themselves are public (https://vesuvius-challenge-open-data.s3.amazonaws.com/<key>).

How the labels were made

The team's ink map of each segment (from its finer scan) is resampled onto the native surface volume's canvas (same registered mesh). Ink = map ≥ 128; supervised = papyrus where the map is ≥ 128 or ≤ 64, so the uncertain band in between is left out of the loss. Every object passed an alignment gate before use (the zero-shot prediction of a 9 µm model must line up with the map within 30 px). Held-out segments (PHerc0009B ×4, PHerc0500P2 ×2, PHerc0343P ×2, PHerc0172 ×2) and every PHerc0841 segment are NOT included: they are the evaluation set.

Limits: the teacher maps are model outputs, not human annotations; letter shapes in them are blobby, and they carry the teacher's errors. PHerc0172's maps come from the same scan as its input, so its labels are easier than the rest.