license: mit
task_categories:
- image-to-text
- text-to-image
language:
- en
pretty_name: simons ARC (abstraction & reasoning corpus) solve symmetry version 23
size_categories:
- 10K<n<100K
configs:
- config_name: default
data_files:
- split: train
path: data.jsonl
Version 1
ARC-AGI Tasks where the job is to transform symmetric images.
example count: 2-4.
test count: 1-2.
image size: 2-3.
symmetry types: hstack2, hstack3, vstack2, vstack3, grid2x2.
Version 2
image size: 2-4.
Version 3
Added HSTACK4, VSTACK4.
Version 4
Added HSTACK5, VSTACK5.
Version 5
Added ImageSymmetrySquare, so images can be rotated by 90 degrees, and flipped over the diagonals.
Version 6
Only exercising ImageSymmetrySquare with HSTACK2, VSTACK2.
Added random padding around the input images, so the job is to extract the content.
Version 7
Only exercising ImageSymmetrySquare with GRID2X2.
Version 8
Only exercising ImageSymmetrySquare with HSTACK3, VSTACK3.
Version 9
Only exercising ImageSymmetrySquare with HSTACK4, HSTACK5, VSTACK4, VSTACK5.
Version 10
Only exercising ImageSymmetrySquare with all the pattern types enabled.
Version 11
Only exercising ImageSymmetrySquare with HSTACK4, VSTACK4.
Version 12
Only exercising ImageSymmetrySquare with HSTACK5, VSTACK5.
Version 13
Only exercising ImageSymmetrySquare with all the pattern types enabled.
Version 14
Enabled ImageSymmetryRect, and added padding.
Enabled ImageSymmetrySquare.
Version 15
Focus only on the new generate_task_with_symmetry_line.
pattern size: 2-4.
padding size: 0-3.
Version 16
Focus only on the new generate_task_with_symmetry_line.
pattern size: 2-5.
padding size: 0-4.
Version 17
Focus only on the new generate_task_with_symmetry_line.
pattern size: 2-5.
padding size: 0-4.
Invert colors of pattern.
Invert colors of background and wall.
Version 18
Focus only on the new generate_task_with_symmetry_line.
pattern size: 2-5.
padding size: 0-4.
Invert colors of pattern.
Invert colors of background and wall.
Sporadic insert decoy pixel inside the pattern that is either background color or wall color.
Random thickness of wall.
wall size: 1-2.
Version 19
Enabled all the task generators.
Slightly smaller patterns for the generate_task_with_symmetry_line, since the big patterns seems to be something that the model struggles grokking.
Version 20
Turned up parameters so the tasks have a length of just below 1024 context length. Previously I aimed for 512 tokens.
Version 21
Focus only on the new generate_task_with_symmetry_line.
pattern size: 2-7.
padding size: 0-10.
wall size: 1-3.
Version 22
Enabled all transformations.
Smaller images, so they can fit inside a context length 1024.
Earlier predictions added to some of the rows.
Version 23
Even smaller images. The model struggles grokking the previous verison.