RAVEN (Hugging Face Datasets version)

RAVEN (Zhang et al., CVPR 2019) converted to the Hugging Face Datasets format for the experiments reported in Staged Self-Supervised Learning for Raven Progressive Matrices (IEEE TNNLS, 2025) and Learning Abstract Visual Reasoning via Task Decomposition: A Case Study in Raven Progressive Matrices (AMCS, 2024). The dataset contains the 70,000 RAVEN tasks with the 6/2/2-fold split used in those experiments: 42,000 train / 14,000 validation / 14,000 test.

Structure

Column	Type	Shape	Description
inputs	uint8	(16, 160, 160)	16 grayscale panels: 8 context (c1–c8) and 8 answer candidates (p1–p8), 160×160 px
target	int8	(16, 113)	ground-truth property vectors for the 16 panels
index	uint8	scalar	index of the correct answer panel

Usage

from datasets import load_dataset

ds = load_dataset("jkwiatkowski/raven")
task = ds["train"][0]

panels = task["inputs"]      # (16, 160, 160), uint8
properties = task["target"]  # (16, 113), int8

Source and preprocessing

• Original RAVEN benchmark: https://github.com/WellyZhang/RAVEN (CVPR 2019).
• Panels are stored at the original 160×160 resolution; in those experiments they are downsampled to 84×84 before being passed to ACT.
• target contains the ground-truth property vectors used to train and evaluate property prediction. Utilities for encoding/decoding property vectors, rendering panels and computing metrics are in raven_tools.

Related resources

• Rule descriptions: https://huggingface.co/datasets/jkwiatkowski/raven_properties
• Model: https://huggingface.co/jkwiatkowski/raven
• Code: https://github.com/jakubkwiatkowski/abstract_compositional_transformer