FaisaI commited on
Commit
0e52f81
·
verified ·
1 Parent(s): 2fe4282
Files changed (1) hide show
  1. README.md +42 -39
README.md CHANGED
@@ -1,39 +1,42 @@
1
- ---
2
- license: cc-by-4.0
3
- pretty_name: VisVerify-Bench (preview)
4
- task_categories:
5
- - image-classification
6
- - object-detection
7
- - image-segmentation
8
- tags:
9
- - annotation-verification
10
- - llm-as-judge
11
- - benchmark
12
- configs:
13
- - config_name: classification
14
- data_files:
15
- - split: dev
16
- path: data/classification/dev-*
17
- - config_name: detection
18
- data_files:
19
- - split: dev
20
- path: data/detection/dev-*
21
- - config_name: segmentation
22
- data_files:
23
- - split: dev
24
- path: data/segmentation/dev-*
25
- ---
26
-
27
- # VisVerify-Bench (preview build)
28
-
29
- **Preview only.** This is the dev split of one domain (everyday scenes, COCO + LVIS),
30
- pushed to check the schema and the dataset viewer. The full card, count tables,
31
- license table and citations come with the full build.
32
-
33
- Each row shows a judge an image plus a *proposed* annotation and stores the
34
- ground truth. `is_correct=true` rows are untouched ground truth (TP). `is_correct=false`
35
- rows were corrupted in a controlled way (`error_type`, `severity`, `iou_with_gt`).
36
-
37
- - `judge_image`: the default rendering shown to a judge (zoomed crop with 50% padding for detection and segmentation; padded crop for classification)
38
- - `image`: the original image (the crop itself for classification)
39
- - `expected`: ground-truth answers for the questions in `questions.json`
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ pretty_name: VisVerify-Bench (preview)
4
+ task_categories:
5
+ - image-classification
6
+ - object-detection
7
+ - image-segmentation
8
+ tags:
9
+ - annotation-verification
10
+ - vlm-as-judge
11
+ - decision-models
12
+ - system-one
13
+ - jev
14
+ - benchmark
15
+ configs:
16
+ - config_name: classification
17
+ data_files:
18
+ - split: dev
19
+ path: data/classification/dev-*
20
+ - config_name: detection
21
+ data_files:
22
+ - split: dev
23
+ path: data/detection/dev-*
24
+ - config_name: segmentation
25
+ data_files:
26
+ - split: dev
27
+ path: data/segmentation/dev-*
28
+ ---
29
+
30
+ # VisVerify-Bench (preview build)
31
+
32
+ **Preview only.** This is the dev split of one domain (everyday scenes, COCO + LVIS),
33
+ pushed to check the schema and the dataset viewer. The full card, count tables,
34
+ license table and citations come with the full build.
35
+
36
+ Each row shows a judge an image plus a *proposed* annotation and stores the
37
+ ground truth. `is_correct=true` rows are untouched ground truth (TP). `is_correct=false`
38
+ rows were corrupted in a controlled way (`error_type`, `severity`, `iou_with_gt`).
39
+
40
+ - `judge_image`: the default rendering shown to a judge (zoomed crop with 50% padding for detection and segmentation; padded crop for classification)
41
+ - `image`: the original image (the crop itself for classification)
42
+ - `expected`: ground-truth answers for the questions in `questions.json`