This repository contains datasets introduced in Hellwig et al. (2026). The datasets are an extension of the letter-string analogies introduced in Lewis & Mitchell (2025).

Each folder contains a dataset with different data attributes, and contains a train, validation, and test set.

The naming convention is:

all_transformations_<copy>_study<N>_perm<N>

all_transformations:

Below are illustrations for each transformation on the standard alphabet.

Transformations present in all splits:
1. Extend Sequence (a b c d -> a b c d e)
2. Successor (a b c d -> a b c e)
3. Predecessor (b c d e -> a c d e)
4. Remove Redundant letter (a b b c d -> a b c d)
5. Fix alphabetic sequence (a b c w e -> a b c d e)
6. Sort (a d c b e -> a b c d e)
7. Sort + Group (aa dd cc bb ee -> aa bb cc dd ee)
8. Remove redundant + Interleave (a x b x b x c x d -> a x b x c x d)
9. Remove Redundant + successor (a b b c d -> a b c e)
10. Fix alphabetic sequence + extend sequence (a b c w -> a b c d e)
Transformations only present in validation/ test splits:
1. Remove Redundant + Sort (a d d c b e -> a b c d e)
2. Extend Sequence + Predecessor (b c d e -> a c d e f)
3. Fix alphabetic sequence + Interleave (a f b f c f w f e -> a f b f c f d f e)
4. Extend sequence + Group (aa bb cc dd -> aa bb cc dd ee)
5. Extend Sequence + Extend Sequence + Successor (a b c d -> a b c d e g)
6. Fix alphabetic Sequence + Predecessor + Successor (a b c w e -> a a c d f)
7. Reverse (a b c d -> d c b a)
8. Shift (a b c d -> e f g h)
9. Replicate (a b c d -> a b c d a b c d)
<copy>: If present, indicates that there are entries in the datasets where study and problem are the same. For datasets with more than one study example, the problem is the same as one of the study examples.
study<N>: How many study examples are presented for each transformation.
perm<N>: How many permuted alphabets were used for generating the given dataset.

Downloads last month: 2

Total file size:

1.47 GB