Experimental datasets

These are experimental datasets, not mature enough for release yet.

Datasets

Easy and Hard curricula.
The filename has the following format: {curriculum name}_{number of samples for each task}x{number of tasks}_{gamma}-{beta}.zip

The easy curriculum is provided with only 100 samples for each of the 20 tasks and it is fully supervised ($\gamma = \beta = 1.0$), the hard curriculum is provided with 100, 1000 and 10000 samples for each of the 18 tasks, with 3 supervision schedules:

• $\gamma = \beta = 1.0$
• $\gamma = 0.8, \beta = 0.2$
• $\gamma = \beta = 0.5$

For the easy curriculum, splits are: $train = 0.5, val = test = 0.25$.
For hard curriculum, splits are: $train = 0.8, val = test = 0.1$.

The ilp_annotations.zip archive contains the train and test annotations for easy and hard curricula with 100 samples to allow symbolic processing without downloading the entire datasets.

The kandy-bongard.zip archive contains the image prompts for visual language model experiments. They are built by taking the first 6 positive (left) and first 6 negative (right) images of the test sets of variants easy_100x20_1.0-1.0 and hard_100x18_1.0-1.0. If less than six images are available for a task.
Images are available in three formats: a simple grid in original KANDY size (224x224 for each panel), and two versions using the Foundalis Bongard Problem style, one in KANDY size and the other in the original Foundalis size (100x100 for each panel). Incomplete samples (due to slight imbalances of the test sets of KANDY) are displayed as white panels.