Wizzas commited on
Commit
ccffaf2
·
verified ·
1 Parent(s): bc6bc66

Plain-language protocol text in the fine-tuned profile manifest

Browse files
results/tuned_deployment_evaluation.json CHANGED
@@ -1,7 +1,7 @@
1
  {
2
  "method": "AnomalyDINO-inspired DINOv2 ViT-S/14 fine-tuned on synthetic defects; 32,768-patch coreset memory",
3
  "profile": "dinov2_tuned_v1",
4
- "protocol": "Official VisA test assignments. Encoder: facebook/dinov2-small with the last 4 blocks fine-tuned on synthetic defects made from the 810 fit normals per category (scripts.encoder_finetune --train --realistic). Memory: a greedy farthest-point coreset of 32,768 patches of all fit-normal patches (the 90 held-out normals are never used). Thresholds: fixed linear quantiles of held-out normal scores (q95; q80 for macaroni1/2). Retrospective development evaluation on test images that were seen during development; not an untouched external test.",
5
  "encoder_sha256": "95ac11984ecae1bb14e9a0d98e85650e73365fc3d57ed49c6bfa526e0498adcb",
6
  "device": "cuda",
7
  "categories": {
 
1
  {
2
  "method": "AnomalyDINO-inspired DINOv2 ViT-S/14 fine-tuned on synthetic defects; 32,768-patch coreset memory",
3
  "profile": "dinov2_tuned_v1",
4
+ "protocol": "Official VisA test assignments. The DINOv2 encoder was fine-tuned on synthetic defects made from normal training images (last 4 blocks only). The memory bank is a 32,768-patch coreset of normal-image patches per product. Thresholds are the 95th percentile of held-out normal scores (80th for Macaroni 1 and 2); held-out normals are never used to build the memory. Retrospective development evaluation on test images that were seen during development; not an untouched external test.",
5
  "encoder_sha256": "95ac11984ecae1bb14e9a0d98e85650e73365fc3d57ed49c6bfa526e0498adcb",
6
  "device": "cuda",
7
  "categories": {