Qualzz20/afv_fiber_9um

Fibre detector fine-tuned on ~9 µm CT

Fine-tuned from ScrollPrize fiber_hz_vt with automatic partial pseudo-labels, centreline/affinity auxiliary losses and geometric compression augmentation. Sources include original CT and native-grid Volcomp q8 CT. Native pitch is 8.64 or 9.362 µm; native grid does not imply uncompressed provenance.

Training lineage: initial → expanded-01 → PHerc0358 full (V3) → multiscroll (V4) → expanded corpus (V5) → centre/void-first augmentation (V6). Published weights are V6 final EMA at step 7624, not the run's best-scoring EMA at 2624.

Outputs retain the base layout: background, vertical, horizontal, intersection. Class 3 received no positive fine-tuning targets and was penalized on known class-0/1/2 voxels; do not rely on it as a validated intersection detector. Only the segmentation backbone is exported.

Training scans span PHerc0125, 0175A, 0191, 0211, 0257, 0268, 0306B, 0358, 0800 and 1447. PHerc0813 and spatially separate PHerc1447 regions provide ancestor-teacher pseudo-label validation, not independent ground truth. No independent test accuracy is claimed.

The validation cohort named real means undeformed: it combines PHerc0813 original CT and PHerc1447 Volcomp q8 CT. Geometric compression augmentation is distinct from the source storage codec.

Inference plans use 128³ windows; loader normalization, precision and mirroring can differ from training. Mirroring is allowed; disabling it trades speed against predictions, without a quantified accuracy claim here.

Training source, recorded configurations and provenance: Qualzz/afv-fiber-9um-training. The archive excludes CT cache and machine orchestration infrastructure; exact reconstruction requires frozen inputs and parent checkpoints that are not bundled.

Matched 3D voxel comparison

Left: original ScrollPrize HZ/VT. Right: the published V6 EMA. Blue denotes vertical fibre predictions; orange denotes horizontal fibre predictions. Both use the same original CT cube, camera, 60% confidence threshold, 128³ inference windows, 64-voxel stride, Gaussian probability blending, BF16 precision and no test-time mirroring. Rendering uses categorical voxels without smoothing or MIP.

The 256³ cube is centred at native XYZ (4032, 4060, 10913) in PHerc0125, at 9.362 µm/voxel. PHerc0125 is present in the training corpus; this is a qualitative illustration, not a held-out accuracy benchmark.

PHerc0125 contributes 25 of the 356 training regions (7.02%) in the frozen final V5/V6 corpus, and 7.863 of 99.526 metres of pseudo-labelled source fibres (7.90%). The V6 sampler selects uniformly among nine training scrolls, so its expected share of sample proposals is 11.11%. These are final-corpus and V6 sampling proportions, not cumulative shares across all fine-tuning phases or base-model pretraining. The displayed cube, including its 32-voxel inference halo, does not overlap any preserved PHerc0125 training-region box in V4–V6. This does not establish independence from base-model pretraining or correlations within the same scroll; the image remains a qualitative comparison, not an independent test. Comparison provenance.