Card: new headline
Browse files
README.md
CHANGED
|
@@ -39,14 +39,14 @@ configs:
|
|
| 39 |
|
| 40 |
# VisVerify-Bench
|
| 41 |
|
| 42 |
-
**
|
| 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)
|
| 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)
|