Team 5
Project Prometheus
Compute where it matters.
Ensemble inference that maps where an ink-detection model can be trusted on an unopened Herculaneum scroll, run as a Fuzzball task array across two nodes.
12.5 min
for an 18-run ensemble
Down from an estimated 2.5 hours on default settings.
1.97×
two-node speedup
98.5% efficiency, with identical outputs.
94.2%
annotator-confirmed ink
Where all 18 runs agree, against 80.2% for a single run.
16 to 20%
of pixels change on retraining
About as much as changing the checkpoint.
Figures are the team's own measurements from the hackathon weekend, as reported in their submission.
The problem
The Vesuvius Challenge reads carbonized Herculaneum scrolls without opening them. A neural network finds faint ink in X-ray CT scans and outputs a probability map. A single map looks equally confident everywhere, and it depends on choices the reader never sees: the checkpoint, the CT layers, the training run. Testing those choices takes many full inferences per segment, which is only routine if the infrastructure makes it cheap.
What they built
A reproducible pipeline that runs the public Vesuvius Challenge ink model, scrollprize/ink_9um, as many independent GPU jobs and then maps where the runs agree. The ensemble covers 3 checkpoints, 3 depth windows and 2 independently trained models on PHerc. 0139.
The output is a triage map. White strokes mark ink in every run and are where to start reading. Red areas, under 0.2% of the papyrus, are deep conflict and form a short list for expert review and new training labels.
How they used the cluster
- Fuzzball workflows stage the model and scans once onto a persistent VAST Data volume, fan out inference as a task array with one NVIDIA GB10 per job on both nodes, then run CPU analysis jobs that read every output from the same volume.
- Tasks skip finished outputs and check that all 17 CT layers were read, so failed runs restart safely.
- With both nodes reading shared storage, mean job time rose 0.5%. Scheduling overhead was 12 to 14 s per phase.
- Packing two jobs per GPU doubled throughput, lifted GPU utilization from 23% to about 60% and cut energy per run by 13 to 22%.
What they found
Consensus is a small core. A single run marks 2.5 to 3.7 times more ink than all nine runs of one model agree on. A different training seed changes 17.6 to 19.5% of pixels, about as much as a different checkpoint and twice as much as a different depth window.
Agreement tracks annotations. Within the labeled part of one segment, precision rose from 80.2% for a single run to 92.1% where one model is unanimous and 94.2% where both are, while recall fell from 64.6% to 34.8%. Rerunning only the most disputed tiles gave 15 to 28% fewer errors than the best uniform subset at the same budget.
Limits the team states
- Agreement is not correctness. Both models share architecture and data.
- The label check is small, and the labels were likely used in training.
- Scaling was not tested past two nodes.