vision/summary.md

Threshold-conversion reproduction: raw summary

• generated: 2026-10-20T21:44:48.114225+00:00
• upstream repo: https://github.com/isllabres/python-multiagent-system
• execution fork: https://github.com/jomasego/python-multiagent-system
• branch: orx/1000-class-vision-vit-b-16-on-cifar-10
• commit: 640f83fde36f5d31a0bcf1ab304bf96c971a22a0
• device: auto (text=cpu, vision=mps if available, digits=cpu)

Theory

Models

model	C	n	argmax acc	delta_0.5	coverage	selective acc	ECE	exact adaptive
1000-class vision: ViT-B/16 (ImageNet head on CIFAR-10)	1000	1000	n/a	51.00%	49.00%	n/a	n/a	True

Temperature sweep

model	T	delta_0.5	argmax acc	ECE
1000-class vision: ViT-B/16 (ImageNet head on CIFAR-10)	0.1	1.00%	n/a	n/a
1000-class vision: ViT-B/16 (ImageNet head on CIFAR-10)	0.25	4.20%	n/a	n/a
1000-class vision: ViT-B/16 (ImageNet head on CIFAR-10)	0.5	12.70%	n/a	n/a
1000-class vision: ViT-B/16 (ImageNet head on CIFAR-10)	1.0	51.00%	n/a	n/a
1000-class vision: ViT-B/16 (ImageNet head on CIFAR-10)	2.0	99.50%	n/a	n/a
1000-class vision: ViT-B/16 (ImageNet head on CIFAR-10)	4.0	100.00%	n/a	n/a
1000-class vision: ViT-B/16 (ImageNet head on CIFAR-10)	8.0	100.00%	n/a	n/a

Per-class ROC operating points (0.5 vs best F1)

model	class	AUC	TPR@0.5	FPR@0.5	best F1 thr	best F1