FaisaI commited on
Commit
ee92d6b
·
verified ·
1 Parent(s): c886077

Card: new headline

Browse files
Files changed (1) hide show
  1. README.md +3 -3
README.md CHANGED
@@ -39,14 +39,14 @@ configs:
39
 
40
  # VisVerify-Bench
41
 
42
- **Can a judge model tell a correct vision annotation from a wrong one?**
43
 
44
  Each item shows a judge an image with a *proposed* class, box or mask. Half the items are
45
  untouched ground truth and should be **accepted**. The other half were corrupted in a
46
  controlled way and should be **rejected**. Every wrong item records *what* is wrong
47
  (`error_type`) and *how badly* (`iou_with_gt`), so accuracy can be plotted against severity.
48
- Built to compare fast "System One" decision models (JEV, Liquid d1-omni) with VLM-as-judge
49
- baselines.
50
 
51
  - 3 tasks (classification, detection, segmentation) x 8 domains (everyday, street, aerial, medical, industrial, agriculture, documents/UI, wildlife/underwater)
52
  - 6,584 items: 5,988 test, 596 dev (separate photos, for prompt tuning only)
 
39
 
40
  # VisVerify-Bench
41
 
42
+ **A benchmark of vision annotation errors for VLM-as-judge and System One vision models.**
43
 
44
  Each item shows a judge an image with a *proposed* class, box or mask. Half the items are
45
  untouched ground truth and should be **accepted**. The other half were corrupted in a
46
  controlled way and should be **rejected**. Every wrong item records *what* is wrong
47
  (`error_type`) and *how badly* (`iou_with_gt`), so accuracy can be plotted against severity.
48
+ Built to compare fast "System One" decision models (JEV, Liquid d1-omni) against
49
+ VLM-as-judge baselines on the same items.
50
 
51
  - 3 tasks (classification, detection, segmentation) x 8 domains (everyday, street, aerial, medical, industrial, agriculture, documents/UI, wildlife/underwater)
52
  - 6,584 items: 5,988 test, 596 dev (separate photos, for prompt tuning only)