metadata

license: mit
task_categories:
  - image-to-text
  - text-to-image
language:
  - en
pretty_name: simons ARC (abstraction & reasoning corpus) combined datasets version 9
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data.jsonl

Version 1

A combination of multiple datasets.

Datasets: dataset_solve_color.jsonl, dataset_solve_rotate.jsonl, dataset_solve_translate.jsonl.

Version 2

Datasets: dataset_solve_color.jsonl, dataset_solve_rotate.jsonl, dataset_solve_translate.jsonl.

Version 3

Datasets: dataset_solve_color.jsonl, dataset_solve_rotate.jsonl, dataset_solve_translate.jsonl.

Version 4

Added a shared dataset name for all these datasets: SIMON-SOLVE-V1. There may be higher version numbers in the future. My hypothesis: Having a version number in the dataset name, it may be easier to unlearn incorrect training data.

Datasets: dataset_solve_color.jsonl, dataset_solve_rotate.jsonl, dataset_solve_translate.jsonl.

Version 5

Different random seed.

Version 6

Using SIMON-SOLVE-V1 everywhere. Remove the SIMON-SOLVE-COLOR, SIMON-SOLVE-ROTATE, SIMON-SOLVE-TRANSLATE.

Version 7

Using SIMON-SOLVE-V1 everywhere.

Version 8

Same settings. Different seed as usual.

Version 9

Switching from context length 256 to context length 512.

Increasing the image sizes so the prompt length stays below 512.

dataset_solve_color, image size: 1-13.

dataset_solve_rotate, image size: 1-9.

dataset_solve_translate, image size: 3-9.