|
--- |
|
license: mit |
|
task_categories: |
|
- image-to-text |
|
- text-to-image |
|
language: |
|
- en |
|
pretty_name: simons ARC (abstraction & reasoning corpus) solve fractal version 8 |
|
size_categories: |
|
- 1K<n<10K |
|
configs: |
|
- config_name: default |
|
data_files: |
|
- split: train |
|
path: data.jsonl |
|
--- |
|
|
|
# Version 1 |
|
|
|
ARC-AGI Tasks where the job is to transform fractal input/output images. |
|
|
|
example count: 2-4. |
|
|
|
test count: 1-2. |
|
|
|
image size: 2-3. |
|
|
|
# Version 2 |
|
|
|
Scale up the input/output images. Scale factor: 1-3. |
|
|
|
# Version 3 |
|
|
|
Scale up the input/output images. Scale factor: 1-3. |
|
|
|
Randomly invert the pattern_image. |
|
|
|
# Version 4 |
|
|
|
Random add padding around the input image, that the model has to crop. |
|
|
|
max_pad_count = 5. |
|
|
|
# Version 5 |
|
|
|
Bigger images |
|
|
|
max_image_size = 5 |
|
|
|
max_pad_count = 8 |
|
|
|
# Version 6 |
|
|
|
Migrating from fitting inside 512 context length to fitting inside 1024 context length. Removing tasks that are too long to fit inside the context length. |
|
|
|
Bigger scale factor: 1-5. |
|
|
|
# Version 7 |
|
|
|
Earlier predictions added to some of the rows. |
|
|
|
Ignoring fractals that RLE compress too badly. Previous cutoff was 1000 tokens. Now it's 800 tokens, so there is room for the earlier prediction. |
|
|
|
# Version 8 |
|
|
|
Added fields: `arc_task`, `test_index`, `earlier_output`. |
