jiahaolu commited on
Commit
f654cf6
·
verified ·
1 Parent(s): fba120a

Describe held-out shapes, message content, and background images

Browse files
Files changed (1) hide show
  1. README.md +7 -0
README.md CHANGED
@@ -76,6 +76,13 @@ field identifies the 75 IID and 75 OOD episodes in each Drag environment.
76
  Rotation and Ten-Choice each use fixed test sets without an IID/OOD subdivision,
77
  so their `distribution` values are `null`.
78
 
 
 
 
 
 
 
 
79
  ## Load episode records
80
 
81
  ```python
 
76
  Rotation and Ten-Choice each use fixed test sets without an IID/OOD subdivision,
77
  so their `distribution` values are `null`.
78
 
79
+ The held-out factors follow the paper's test-set construction. Drag's IID
80
+ episodes use training shape categories, while its OOD episodes use held-out
81
+ shape categories. All Ten-Choice test scenes draw from 240 messages disjoint
82
+ from the 80 training messages. Rotation uses 150 test background images
83
+ disjoint from the 1,000 training backgrounds. Ten-Choice and Rotation therefore
84
+ evaluate held-out content across their full test sets.
85
+
86
  ## Load episode records
87
 
88
  ```python