mlfoundations-cua-dev/easyr1-10k-hard-qwen7b-easy-gta1-composed-20-dual-5-montage-4MP

easyr1-10k-hard-qwen7b-easy-gta1-composed-20-dual-5-montage-4MP

This dataset was generated using the enhanced EasyR1 grounding dataset pipeline with composition capabilities.

Generation Details

• —Generated on: 2025-08-24 11:02:10 UTC
• —Script: push_easyr1_composed_to_hf.py
• —Data directory: /lustre/fs12/portfolios/nvr/projects/nvr_lacr_llm/users/aawadalla/LLaMA-Factory/data

Parameters Used

• —Maximum samples: 10000
• —Image resize (max megapixels): 4.0 MP
• —Minimum native image resolution: 0.0 MP
• —Prompt format: gta1_with_resolution
• —Output format: coordinates
• —Random seed: 42

Composition Settings

• —Dual-screen ratio: 0.2 (20% of samples)
• —Montage ratio: 0.05 (5% of samples)
• —Single-screen ratio: 0.75 (75% of samples)

Dataset Groups

The following JSON/JSONL files were used to create this dataset:

Dataset Group 1

Files (intersection of kept samples across all files):

• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/pixmo-points-qwentoolcall-notgrounded-QwenQwen2.5-VL-7B-Instruct-qwentoolcall.jsonl
• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/pixmo-points-gta1-correctlygrounded-HelloKKMeGTA1-7B-gta1.jsonl

Dataset Group 2

Files (intersection of kept samples across all files):

• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/autogui-grounding-only-qwentoolcall-notgrounded-QwenQwen2.5-VL-7B-Instruct-qwentoolcall.jsonl
• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/autogui-grounding-only-gta1-correctlygrounded-HelloKKMeGTA1-7B-gta1.jsonl

Dataset Group 3

Files (intersection of kept samples across all files):

• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/seeclick-qwentoolcall-notgrounded-QwenQwen2.5-VL-7B-Instruct-qwentoolcall.jsonl
• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/seeclick-gta1-correctlygrounded-HelloKKMeGTA1-7B-gta1.jsonl

Dataset Group 4

Files (intersection of kept samples across all files):

• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/pc-e-grounding-only-claude-instructions-qwentoolcall-notgrounded-QwenQwen2.5-VL-7B-Instruct-qwentoolcall.jsonl
• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/pc-e-grounding-only-claude-instructions-gta1-correctlygrounded-HelloKKMeGTA1-7B-gta1.jsonl

Dataset Group 5

Files (intersection of kept samples across all files):

• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/omniact-grounding-only-qwentoolcall-notgrounded-QwenQwen2.5-VL-7B-Instruct-qwentoolcall.jsonl
• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/omniact-grounding-only-gta1-correctlygrounded-HelloKKMeGTA1-7B-gta1.jsonl

Dataset Group 6

Files (intersection of kept samples across all files):

• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/showui-desktop-qwentoolcall-notgrounded-QwenQwen2.5-VL-7B-Instruct-qwentoolcall.jsonl
• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/showui-desktop-gta1-correctlygrounded-HelloKKMeGTA1-7B-gta1.jsonl

Dataset Group 7

Files (intersection of kept samples across all files):

• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/showui-web-qwentoolcall-notgrounded-QwenQwen2.5-VL-7B-Instruct-qwentoolcall.jsonl
• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/showui-web-gta1-correctlygrounded-HelloKKMeGTA1-7B-gta1.jsonl

Dataset Group 8

Files (intersection of kept samples across all files):

• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/uground-qwentoolcall-notgrounded-QwenQwen2.5-VL-7B-Instruct-qwentoolcall.jsonl
• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/uground-gta1-correctlygrounded-HelloKKMeGTA1-7B-gta1.jsonl

Dataset Group 9

Files (intersection of kept samples across all files):

• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/waveui-qwentoolcall-notgrounded-QwenQwen2.5-VL-7B-Instruct-qwentoolcall.jsonl
• —/lustre/fs12/portfolios/nvr/projects/nvrlacrllm/users/aawadalla/grounding-data-filters/waveui-gta1-correctlygrounded-HelloKKMeGTA1-7B-gta1.jsonl

Dataset Statistics

• —Total training samples: 10000
• —Composition breakdown:
• —Single-screen: 7500
• —Dual-screen: 2000
• —Montage: 500
• —Image dimensions: Variable (based on composition type)
• —Columns: images, prompt, easyr1prompt, bbox, imagepath, messages, imagewidth, imageheight, compositiontype

System Prompt

The following system prompt is used for this dataset:

You are an expert UI element locator. Given a GUI image and a user's element description, provide the coordinates of the specified element as a single (x,y) point. The image resolution is height 2048 and width 2048. For elements with area, return the center point.

Output the coordinate pair exactly:
(x,y)

Sample Entry

• —User prompt: <image> Find and click Sport
• —Assistant response: (891,172)
• —Bounding box: [865, 159, 917, 185]
• —Image path: showui-web-images/showuiweb021834.jpg
• —Composition type: single

Usage

from datasets import load_dataset

dataset = load_dataset("mlfoundations-cua-dev/easyr1-10k-hard-qwen7b-easy-gta1-composed-20-dual-5-montage-4MP")

# Access the training data
train_data = dataset['train']

# Example: Get samples by composition type
single_screen = [s for s in train_data if s.get('_composition_type') == 'single']
dual_screen = [s for s in train_data if s.get('_composition_type') == 'dual_screen']
montage = [s for s in train_data if s.get('_composition_type') == 'montage']

Composition Types

Single-Screen

Standard single image samples with UI element grounding.

Dual-Screen

Two images concatenated horizontally, simulating dual-monitor setups. Coordinates are adjusted to the correct screen half.

Montage

Multiple application windows overlaid on desktop backgrounds, simulating real multi-window desktop environments.

License

Please refer to the original dataset licenses for usage restrictions.