Micasa997/two-box-judge-gui-sharded

Two-Box Judge GUI Dataset (Sharded)

A multimodal dataset for training GUI element selection models, packaged in WebDataset format for efficient streaming.

Dataset Statistics

Split	Samples	Shards	Size
Train	115,638	6	25.32 GB
Validation	12,849	1	2.82 GB

Format

This dataset uses WebDataset format - sharded tar.gz archives for efficient streaming:

train/
├── shard-00000.tar.gz
├── shard-00001.tar.gz
└── ...

Each shard contains multiple samples in this structure:

000000.image1.png   # GUI with green box (candidate 1)
000000.image2.png   # GUI with red box (candidate 2)
000000.json         # {"sample_id": 0, "user_command": "...", "label": "1"}
000001.image1.png
000001.image2.png
000001.json
...

Usage

With WebDataset Library

import webdataset as wds

dataset = wds.WebDataset("train/shard-{00000..00115}.tar.gz")
    .decode("pil")
    .to_tuple("image1.png", "image2.png", "json")

for image1, image2, metadata in dataset:
    print(metadata["user_command"], metadata["label"])

Manual Loading

import tarfile
import json
from PIL import Image
import io

with tarfile.open("train/shard-00000.tar.gz", "r:gz") as tar:
    for member in tar.getmembers():
        if member.name.endswith(".json"):
            f = tar.extractfile(member)
            metadata = json.load(f)
            print(metadata)

Task Description

Given two images showing GUI elements with bounding boxes:

• Image 1: Green box labeled "1" (first grounding result)
• Image 2: Red box labeled "2" (re-grounding result after masking)

The model predicts which box better fulfills the user's command.

License

Apache 2.0