ReGuLaR Data

This repository contains the ReGuLaR training mix with images, question-answer pairs, bounding boxes, and relation annotations. Metadata is provided as parquet shards; images are provided as compressed source-specific archives to keep the repository uploadable.

Files

data/train-xxxxx-of-yyyyy.parquet
image_archives/*.tar.zst
records.jsonl
manifest.json
stats.json
extract_images.sh

The parquet schema contains:

• id, source, dataset
• image: primary image path, relative to the dataset root after extraction
• images: all image paths for the sample; mGrounding samples may contain multiple images
• question, answer
• relations_json: subject-relation-object annotations serialized as JSON
• scene_graph_json: compact scene graph or relation evidence serialized as JSON
• metadata_json: source-specific metadata serialized as JSON

Download And Extract Images

from huggingface_hub import snapshot_download
snapshot_download("zihuwang/ReGuLaR", repo_type="dataset", local_dir="regular_data")

Then extract the image archives from the dataset root:

cd regular_data
bash extract_images.sh

Equivalently, you can extract archives manually:

for f in image_archives/*.tar.zst; do tar --zstd -xf "$f"; done

This creates the images/ directory expected by the metadata paths:

images/gqa/...
images/openimage/...
images/clevr/...
images/visual_genome/...
images/psg/...
images/mgrounding/...
images/vsr/...

Usage

import json
from pathlib import Path
from PIL import Image
from datasets import load_dataset

root = Path("regular_data")
ds = load_dataset("parquet", data_files=str(root / "data/*.parquet"), split="train")

ex = ds[0]
image = Image.open(root / ex["image"]).convert("RGB")
all_images = [Image.open(root / p).convert("RGB") for p in ex["images"]]

relations = json.loads(ex["relations_json"])
scene_graph = json.loads(ex["scene_graph_json"])
metadata = json.loads(ex["metadata_json"])

Each relation follows a subject-relation-object format. Subject and object entries include names and bounding boxes when available. Bounding boxes use bbox_xyxy.

Notes

The dataset combines examples derived from GQA, OpenImage, CLEVR, Visual Genome, PSG, mGrounding, and VSR. Please follow the licenses and terms of the original data sources.