BOLD5000-QA

An fMRI question-answering dataset built on top of the BOLD5000 fMRI dataset. BOLD5000-QA pairs fMRI recordings of subjects viewing natural images with compositional question-answer pairs derived from scene graphs.

This dataset is introduced in Neuro-Symbolic Decoding of Neural Activity (ICLR 2026).

Dataset Description

BOLD5000-QA converts visual scene graphs from BOLD5000 images into structured QA pairs. Each sample contains:

• fMRI data: Voxel-level brain activity recorded while a subject views an image, parcellated by brain atlas (e.g., Yeo-17 networks)
• Queries: Symbolic compositional queries about the image content (e.g., scene() -> filter(person) -> query(holding, ?))
• Answers: Ground-truth answers (yes/no for Boolean queries, or vocabulary tokens for attribute queries)

Statistics

	Training	Test
QA examples	~133K	~2K
Subjects	4	4

Subjects

• CSI1, CSI2, CSI3, CSI4

Dataset Structure

BOLD5000-QA/
  <subject>/          # e.g., CSI1
    train/
      <img_id>.npy    # Per-image data (queries, answers, brain_region)
    test/
      <img_id>.npy

Each .npy file is a dictionary containing:

• queries (list of str): Symbolic query programs
• answers (list of str): Corresponding answers
• brain_region (np.ndarray): fMRI activation parcellated by atlas

Usage

import numpy as np

sample = np.load("BOLD5000-QA/CSI1/train/0.npy", allow_pickle=True).item()
print(sample['queries'])       # list of symbolic query strings
print(sample['answers'])       # list of answer strings
print(sample['brain_region'].shape)  # fMRI region activations

Or use the provided PyTorch dataset loader from the NEURONA codebase:

from loader.fqa import FQADataset

dataset = FQADataset(data_dir="data/BOLD5000-QA", split="train", subject="CSI1")

Links

• Paper: Neuro-Symbolic Decoding of Neural Activity
• Code: github.com/PPWangyc/neurona
• Project Page: ppwangyc.github.io/projects/neurona

Citation

@article{wang2026neuro,
  title={Neuro-Symbolic Decoding of Neural Activity},
  author={Wang, Yanchen and Hsu, Joy and Adeli, Ehsan and Wu, Jiajun},
  journal={arXiv preprint arXiv:2603.03343},
  year={2026}
}