README.md

metadata

language:
  - zh
  - en
license: cc-by-nc-4.0
size_categories:
  - 1K<n<10K
task_categories:
  - visual-question-answering
  - image-text-to-text
tags:
  - gui-agent
  - benchmark
  - multimodal
pretty_name: UniGUI-Bench
configs:
  - config_name: element_grounding
    data_files:
      - split: test
        path: data/process_eval/element_grounding.json
    field: examples
  - config_name: action_understanding
    data_files:
      - split: test
        path: data/process_eval/action_understanding.json
    field: examples
  - config_name: page_understanding
    data_files:
      - split: test
        path: data/process_eval/page_understanding.json
    field: examples
  - config_name: action_anticipation
    data_files:
      - split: test
        path: data/process_eval/action_anticipation.json
    field: examples
  - config_name: intent_reasoning
    data_files:
      - split: test
        path: data/process_eval/intent_reasoning.json
    field: examples
  - config_name: action_execution
    data_files:
      - split: test
        path: data/process_eval/action_execution.json
    field: examples
  - config_name: dynamic_adaptation
    data_files:
      - split: test
        path: data/process_eval/dynamic_adaptation.json
    field: examples
  - config_name: task_planning
    data_files:
      - split: test
        path: data/process_eval/task_planning.json
    field: examples
  - config_name: task_progress_awareness
    data_files:
      - split: test
        path: data/process_eval/task_progress_awareness.json
    field: examples
  - config_name: task_completion_detection
    data_files:
      - split: test
        path: data/process_eval/task_completion_detection.json
    field: examples
  - config_name: intern_state_memory
    data_files:
      - split: test
        path: data/process_eval/intern_state_memory.json
    field: examples

UniGUI-Bench

A comprehensive benchmark for evaluating GUI agents across 11 capability dimensions.

Dataset Structure

System Evaluation

• trajectory: 1,200 complete GUI task trajectories with step-by-step screenshots
• gui_system: 7,108 individual step evaluation samples

Process Evaluation (11 Abilities, 200 samples each)

Ability	Description
element_grounding	Locating UI elements given instructions
action_understanding	Understanding what action was performed
page_understanding	Comprehending page content and layout
action_anticipation	Predicting the next action
intent_reasoning	Inferring user intent from actions
action_execution	Executing the correct action
dynamic_adaptation	Adapting to changing interfaces
task_planning	Planning multi-step tasks
task_progress_awareness	Tracking task completion progress
task_completion_detection	Detecting when a task is done
intern_state_memory	Remembering internal states

File Structure

├── data/
│   ├── system_eval/
│   │   ├── trajectory.json
│   │   └── gui_system.json
│   └── process_eval/
│       ├── element_grounding.json
│       ├── action_understanding.json
│       └── ... (11 ability files)
├── screenshots/
│   ├── tasks/{task_id}/step_XX.png
│   └── abilities/{ability_name}/
└── manifest.json

Usage

from datasets import load_dataset

# Load a specific ability
ds = load_dataset("AGI-Eval/UniGUI-Bench", "element_grounding")

# Access samples
for sample in ds["test"]:
    print(sample["input"])
    print(sample["target"])
    print(sample["extra_data"])