README.md

metadata

license: gpl-3.0
task_categories:
  - visual-question-answering
  - question-answering
language:
  - en
tags:
  - abstract-reasoning
  - cognitive-evaluation
  - raven-progressive-matrices
  - neuropsychology
  - llm-evaluation
pretty_name: NeuroCognition RAPM
size_categories:
  - n<1K

NeuroCognition RAPM Dataset

Evaluation dataset for the paper "A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities" (arXiv:2603.02540), submitted to NeurIPS 2026 Evaluations & Datasets Track.

This dataset contains two variants of Raven's Progressive Matrices (RAPM) used to evaluate abstract reasoning in large language models:

Dataset Structure

Visual RAPM (visual_rapm.json + images/)

• 140 items drawn from the RAVEN dataset (Zhang et al., CVPR 2019)
• Each item is a composite image combining the 3×3 problem matrix and 8 answer choices into a single figure, rendered from raw RAVEN .npz files using matplotlib
• Covers 5 RAVEN dataset types: center_single, distribute_four, distribute_nine, in_center_single_out_center_single, left_center_single_right_center_single, up_center_single_down_center_single

Fields:

Field	Description
id	Unique identifier (e.g. RAVEN_238_test_center_single)
dataset_type	RAVEN configuration type
full_image	Path to composite image (problem matrix + answer choices)
correct_answer	Index of the correct answer (0-based, out of 8 choices)

Text RAPM (text_rapm.jsonl)

• 200 items of an original text-based analogue of RAPM
• Each item presents a 3×3 matrix of character strings instead of geometric figures, where cells follow quantitative or structural rules across rows and columns
• Minimum difficulty of 20 (axis attempts) and designed to require multi-step rule inference

Fields:

Field	Description
question_grid	3×3 list of strings (last cell is null)
options	8 candidate answers
correct_index	Index of the correct answer
answer	Correct answer string
row_attribute	Rule governing rows
col_attribute	Rule governing columns
cell_constraints	Per-cell character constraints
axis_attempts	Difficulty indicator (generation attempts needed)

Usage

import json

# Visual RAPM
with open("visual_rapm.json") as f:
    visual_data = json.load(f)["questions"]

# Text RAPM
with open("text_rapm.jsonl") as f:
    text_data = [json.loads(line) for line in f]

Scripts

scripts/render_raven_images.py contains the script used to render composite images from raw RAVEN .npz files. Raw .npz files are available from the original RAVEN repository.

License & Attribution

Visual RAPM images are derived from the RAVEN dataset and are distributed under the same GPL-3.0 license. Please cite the original RAVEN paper:

@inproceedings{zhang2019raven,
  title={RAVEN: A Dataset for Relational and Analogical Visual rEasoNing},
  author={Zhang, Chi and Gao, Feng and Jia, Baoxiong and Zhu, Yixin and Zhu, Song-Chun},
  booktitle={CVPR},
  year={2019}
}

Text RAPM is original work by the authors of this paper.

If you use this dataset, please also cite our paper:

@article{haznitrama2026neurocognition,
  title={A Neuropsychologically Grounded Evaluation of LLM Cognitive Abilities},
  author={Haznitrama, Faiz Ghifari and Ardi, Faeyza Rishad and Oh, Alice},
  journal={arXiv:2603.02540},
  year={2026}
}

RAI Metadata

• Intended use: Evaluation of abstract reasoning in LLMs; research only
• Out-of-scope use: Training data for commercial models without respecting GPL-3.0 terms
• Data limitations: Visual RAPM items are a subset (140/10000) of RAVEN test split; text RAPM uses ASCII-like character strings which may not fully capture visuospatial reasoning
• Sensitive information: None
• Social impact: Contributes to understanding of cognitive capabilities and limitations of LLMs, grounded in neuropsychological theory
• Provenance: Visual RAPM rendered from RAVEN-10000 (Zhang et al., CVPR 2019); Text RAPM generated by the authors