ilp-datasets / iparc /README.md
andrewcropper's picture
Add files using upload-large-folder tool
1dbb54d verified
|
Raw
History Blame Contribute Delete
1.9 kB

IPARC

ILP-flavoured ARC (Abstraction and Reasoning Corpus) tasks, encoded as ILP problems. Each task requires learning a program that transforms a 2D grid of coloured cells. Tasks are grouped into six categories by difficulty and transformation type. 600 tasks total (100 per category).

Tasks: 6 categories × 100 tasks = 600 tasks (in <Category>/Task<NNN>/train/)

Category Description
CatA_Simple Category A transformations, simple difficulty
CatA_Hard Category A transformations, hard difficulty
CatB_Selection Category B: select a subset of input cells for the output
CatB_Iteration Category B: iterate over rows or columns to compute output
CatB_Sequence Category B: sequential (multi-step) transformations
CatB_Hard Category B transformations, hard difficulty

Head relation: out(Ex, Row, Col, Value) — cell (Row, Col) in the output grid of example Ex has colour Value.

Background relations:

Relation Arity Description
in(Ex, Row, Col, Value) 4 Input grid: cell (Row, Col) in example Ex has colour Value
empty(Ex, Row, Col) 3 Cell (Row, Col) is empty (colour 0) in the input of example Ex
width(Ex, W) 2 Input grid of example Ex has width W
height(Ex, H) 2 Input grid of example Ex has height H
add(V1, V2, V3) 3 V3 = V1 + V2 (integer arithmetic)
lt(V1, V2) 2 V1 < V2
my_succ(V1, V2) 2 V2 = V1 + 1
position_N/1 1 Type-constant: value is position N (1–15)
value_N/1 1 Type-constant: value is colour/value N

Source: Challa, A., Srinivasan, A., Bain, M., & Shroff, G. "A Program-Synthesis Challenge for ARC-Like Tasks." In Muggleton, S. & Tamaddoni-Nezhad, A. (eds.), ILP 2022, LNCS vol. 13779, pp. 25–39. Springer, 2022.