README.md

metadata

configs:
  - config_name: close_ended
    data_files:
      - split: val
        path: close_ended.json
  - config_name: open_ended
    data_files:
      - split: val
        path: open_ended.json
license: mit
task_categories:
  - visual-question-answering
language:
  - en
tags:
  - spatial-reasoning
  - object-location
  - direction

Object Location 4-Class (Subset 1 of 5)

A visual question answering dataset for evaluating spatial reasoning — specifically, identifying the location of an object (Up / Down / Left / Right) from the viewer's perspective.

Dataset Configurations

close_ended

Multiple-choice format with 4 fixed options (A: Up, B: Down, C: Left, D: Right).

open_ended

Free-form answer format. The model is expected to respond with one of: Up, Down, Left, Right.

Usage

from datasets import load_dataset

# Close-ended (multiple choice)
ds_close = load_dataset("YOUR_HF_ID/object_location_4class", name="close_ended", split="val")

# Open-ended (free-form)
ds_open = load_dataset("YOUR_HF_ID/object_location_4class", name="open_ended", split="val")

Data Fields

Close-ended

Field	Type	Description
id	int	Sample index
image	string	Image filename (e.g., rectangle_yellow_123.png)
question	string	Question text
options	list[string]	["Up", "Down", "Left", "Right"]
answer	string	Correct option letter (A/B/C/D)
answer_text	string	Correct answer in text (Up/Down/Left/Right)
category	string	Ground truth direction folder (up/down/left/right)

Open-ended

Field	Type	Description
id	int	Sample index
image	string	Image filename (e.g., rectangle_yellow_123.png)
question	string	Question text
answer	string	Correct answer (Up/Down/Left/Right)
category	string	Ground truth direction folder (up/down/left/right)

Notes

• Images are not included in this repository. They should be stored locally and loaded via the category and image fields to reconstruct the full path:

  {DATA_ROOT}/{category}/{image}