neoneye's picture
Update README.md
fa647ba verified
metadata
license: mit
task_categories:
  - image-to-text
  - text-to-image
language:
  - en
pretty_name: simons ARC (abstraction & reasoning corpus) solve mass version 12
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 identify the mass of objects with a specific size.

example count: 2-4.

test count: 1-2.

image size: 4-6.

find mass: 1-2.

connectivity: ALL8.

Version 2

image size: 3-10.

Version 3

image size: 1-15.

Version 4

image size: 1-15.

find mass: 1-3.

Version 5

image size: 1-8.

find mass: 1-4.

Version 6

Compare mass of adjacent rows/columns. image size: 4-7. color count: 10.

This was something that the model struggles with. What if I use fewer colors and smaller images.

Version 7

Compare mass of adjacent rows/columns. image size: 4-6. color count: 2,3,10.

Still something that the model has difficulties with.

Version 8

Disabled comparing columns. Disable too many colors.

Focus is only on compare mass of adjacent rows. image size: 4-6. color count: 2.

Yay, the model groks this.

Version 9

Disabled comparing columns. Disable too many colors.

Focus is only on compare mass of adjacent columns. image size: 4-6. color count: 2.

This is hard for the model to make sense of.

Version 10

Still focus is only on compare mass of adjacent columns. image size: 3-5. color count: 2.

Version 11

Focus on adjacent rows/columns. image size: 3-8. color count: 2,3,10.

Version 12

Bigger images. image size: 3-12.