Append 20,000 vision training rows; preserve previous data and evaluation

CHANGELOG.md

@@ -4,6 +4,11 @@ Changes that affect published data, newest first: one entry per publish, `## <da
 concise bullets on what changed in the data. `tasksource-jev-typed-decisions` releases are tagged on the Hub
 (`load_dataset("tasksource/tasksource-jev-typed-decisions", revision="v2026.10.13")`).
 
+## tasksource-jev-typed-decisions v2026.10.17
+
+- `vision`: add ShapeWorld object-box grounding, capped at 20k training rows from the 100k-QA procedural mirror. Native scene models verify every unique target and spatial relation.
+- Preserve existing training and evaluation shards and exclude images already present in published evaluation. Keep generator seeds, geometry and original MIT license evidence.
+
 ## tasksource-jev-typed-decisions v2026.10.16.2
 
 - `vision`: add Flickr30k Entities phrase grounding, capped at 20k training rows from the 100k-QA mirror. Full captions identify the queried phrase; options are native annotated regions. Published 20,000 new rows, bringing vision training to 702,314 rows.

README.md

@@ -207,16 +207,16 @@ dataset_info:
     dtype: string
   splits:
   - name: train
-    num_bytes: 236748770808
-    num_examples: 702314
+    num_bytes: 236883226428
+    num_examples: 722314
   - name: test
     num_bytes: 146575133
     num_examples: 475
   - name: validation
     num_bytes: 274444703
     num_examples: 972
-  download_size: 140047455255
-  dataset_size: 237169790644
+  download_size: 140122509371
+  dataset_size: 237304246264
 ---
 
 # tasksource-jev-typed-decisions

vision/release-audit.json

@@ -3,7 +3,7 @@
   "config": "vision",
   "splits": {
     "train": {
-      "rows": 702314,
+      "rows": 722314,
       "sources": {
         "vision/tallyqa/count": 30000,
         "vision/nlvr2": 30000,
@@ -54,7 +54,8 @@
         "vision/refcoco/box": 6665,
         "vision/refcoco/plus/box": 6650,
         "vision/refcoco/g/box": 6685,
-        "vision/flickr30k/box": 20000
+        "vision/flickr30k/box": 20000,
+        "vision/shapeworld/box": 20000
       },
       "families": {
         "vision/tallyqa": 30000,
@@ -88,12 +89,13 @@
         "vision/mind2web": 15258,
         "vision/spatial-dise": 2410,
         "vision/refcoco": 20000,
-        "vision/flickr30k": 20000
+        "vision/flickr30k": 20000,
+        "vision/shapeworld": 20000
       },
       "license_use": {
         "unspecified": 171691,
         "non-commercial": 156616,
-        "commercial": 374007
+        "commercial": 394007
       }
     },
     "test": {
@@ -453,6 +455,10 @@
     {
       "manifest": "vision/additions/82c3f66baf41/manifest.json",
       "train_rows": 20000
+    },
+    {
+      "manifest": "vision/additions/28a7c1167437/manifest.json",
+      "train_rows": 20000
     }
   ],
   "removals": [

vision/sources.yaml

@@ -33,6 +33,7 @@ datasets:
 - tasksource/mentis-oculi-paper-fold
 - tasksource/multimodal-mind2web
 - tasksource/refcoco-grounding
+- tasksource/shapeworld-grounding
 - tasksource/spair71k-grid
 - tasksource/spatial-dise
 - tasksource/sphinx-mcq
@@ -978,4 +979,23 @@ sources:
       scope: Human phrase/box annotations; photograph copyrights remain with their
         owners, not the dataset publishers.
       url: https://github.com/BryanPlummer/flickr30k_entities#flickr30k-entities-dataset
+  vision/shapeworld/box:
+    rows:
+      train: 20000
+    dataset: tasksource/shapeworld-grounding
+    revision: fe3756236e56d627072f4809a04ee962f4b92af4
+    revisions:
+      tasksource/shapeworld-grounding: fe3756236e56d627072f4809a04ee962f4b92af4
+    license: mit, mit (original source)
+    license_use: commercial
+    card_licenses:
+      tasksource/shapeworld-grounding:
+      - mit
+    source_license_evidence:
+      license: mit
+      license_use: commercial
+      checked: '2026-10-17'
+      scope: Native ShapeWorld renderer and generated geometric scenes; decisions
+        verified on saved scene models.
+      url: https://github.com/AlexKuhnle/ShapeWorld/blob/e720bf46e57fc01326d04d639fa6133d9c12158f/LICENSE
 parent_tasksource_commit: f9962e198e653bfb73cc2ae5ea33e1511b6cb746

vision/train-append-28a7c1167437-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99ab2a769cec3f41f1ad2e17c9df4bce227b65cd46139d1a86e6770126f75f50
+size 75054116