OmniAgentBench Dataset Card

Overview

OmniAgentBench is a comprehensive benchmark for evaluating multimodal agents under realistic "wild" conditions including speech input, acoustic noise, and complex multi-turn interactions.

Organization: omniagentbench
Dataset: OmniAgentBench
Total Size: 27.1 GB
Contributors: Hodfa71, acbueff

---

Dataset Structure

The dataset is organized into separate benchmark folders at the root level:

OmniAgentBench/
├── 📁 data/                      # Dataset metadata and manifests
├── 📁 dataset/mpcc/              # MPCC dataset files (tabular)
├── 📁 gui_odyssey/               # ⭐ GUI Odyssey (NEW - 1,800 wild audio samples)
│   ├── General_Tool/
│   ├── Information_Management/
│   ├── Multi_Apps/
│   ├── Media_Entertainment/
│   ├── Social_Sharing/
│   ├── Web_Shopping/
│   └── screenshots/              # GUI Odyssey screenshots ONLY
├── 📁 images/mpcc/               # ⚠️ MPCC screenshots ONLY (5,700 files)
├── 📁 mpcc/                      # MPCC audio and manifests
│   ├── manifest.json
│   └── *.wav (2,700+ audio files)
├── 📁 wild/                      # Shared noise resources (MUSAN, etc.)
└── 📁 wild_long_scattered/       # Additional wild variations

---

⚠️ Important: Folder Separation

The images/ folder contains ONLY MPCC screenshots!

Folder	Benchmark	Content
images/mpcc/	MPCC	5,700 screenshots (flight schedules, calendars, meetings)
gui_odyssey/screenshots/	GUI Odyssey	1,950 mobile app screenshots
gui_odyssey/General_Tool/screenshots/	GUI Odyssey	Per-category screenshots

They are NOT mixed - each benchmark has its own dedicated folder.

---

Benchmarks

1. MPCC (Multi-Modal Planning and Control Challenge)

Constraint planning over visual schedules (flights, calendars, meetings).

Location: mpcc/, images/mpcc/, dataset/mpcc/

Contents:

• Audio: 300 speech samples across 3 tasks × 3 difficulties
• Images: 5,700 screenshots (schedules, flight info, calendars)
• Text: Structured task instructions with JSON output format
• Ground Truth: Optimal plans with constraint satisfaction labels

Format:

{
  "audio_file": "mpcc_flight_easy_1.wav",
  "image_paths": ["images/mpcc/flight_easy_1_img1.jpg"],
  "text_instruction": "Find flights from London to Vienna...",
  "gold_answer": {"flight_way": "...", "price": 291}
}

---

2. GUI Odyssey (NEW - Wild Audio)

Cross-app mobile GUI navigation with wild audio inputs.

Location: gui_odyssey/

Contents:

• Audio: 1,800+ wild samples (300 per category × 6 categories)
  • TTS-generated with Qwen3-TTS
  • Acoustic noise: coffee_shop, convention_hall, outdoor, etc.
  • SNR: 5dB, 10dB, 15dB
• Images: 1,950 mobile app screenshots
• Text: Task instructions (spoken in audio)
• Ground Truth: Click coordinates, text inputs, scroll actions

Categories:

• General_Tool (300 samples)
• Information_Management (300 samples)
• Web_Shopping (400 samples)
• Multi_Apps (300 samples)
• Media_Entertainment (400 samples)
• Social_Sharing (300 samples)

Format per sample:

{
  "sample_id": "0182869798349621_step2",
  "instruction": "Use Bloomberg to search for Pfizer stock news...",
  "audio_path": "gui_odyssey/General_Tool/audio/0182869798349621_step2_summer_outdoor_snr15.wav",
  "screenshot_path": "gui_odyssey/screenshots/General_Tool/0182869798349621_2.png",
  "gt_action": "CLICK",
  "gt_x": 351.0,
  "gt_y": 54.0,
  "gt_all_steps": "[...full episode with 11 steps...]"  # JSON string
}

---

Usage Examples

Load MPCC

from datasets import load_dataset

# Load MPCC manifest
import json
with open("mpcc/manifest.json") as f:
    mpcc_data = json.load(f)

# Access sample
sample = mpcc_data["mpcc_flight_easy_1"]
audio = sample["audio_file"]  # Path to .wav
images = sample["image_paths"]  # List of image paths

Load GUI Odyssey

import pandas as pd
from huggingface_hub import hf_hub_download

# Download parquet
parquet_path = hf_hub_download(
    repo_id="omniagentbench/OmniAgentBench",
    filename="gui_odyssey/General_Tool/gui_odyssey_General_Tool_wild.parquet",
    repo_type="dataset"
)

# Load
df = pd.read_parquet(parquet_path)

# Access sample
row = df.iloc[0]
print(row["instruction"])  # Text instruction
print(row["audio_path"])   # Path to audio
print(row["screenshot_path"])  # Path to screenshot
print(row["gt_action"])    # Ground truth: CLICK/TEXT/SCROLL

Access Audio and Images

from huggingface_hub import hf_hub_download

# MPCC audio
audio_path = hf_hub_download(
    repo_id="omniagentbench/OmniAgentBench",
    filename="mpcc/mpcc_flight_easy_1.wav",
    repo_type="dataset"
)

# MPCC image
image_path = hf_hub_download(
    repo_id="omniagentbench/OmniAgentBench",
    filename="images/mpcc/calendar_easy_0_img1.jpg",
    repo_type="dataset"
)

# GUI Odyssey audio
audio_path = hf_hub_download(
    repo_id="omniagentbench/OmniAgentBench",
    filename="gui_odyssey/General_Tool/audio/0182869798349621_step0_coffee_shop_snr5.wav",
    repo_type="dataset"
)

# GUI Odyssey screenshot
screenshot_path = hf_hub_download(
    repo_id="omniagentbench/OmniAgentBench",
    filename="gui_odyssey/screenshots/General_Tool/0182869798349621_0.png",
    repo_type="dataset"
)

---

Statistics

Benchmark	Audio Files	Images	Size	Samples
MPCC	2,700+	5,700	~10 GB	300 speech
GUI Odyssey	1,900+	1,950	~3 GB	1,800 wild
Total	4,600+	7,650	~27 GB	2,100+

---

Ground Truth Format

MPCC

• Task Type: Constraint planning
• Output: JSON with flight_way/schedule + price
• Evaluation: Feasible rate, optimal rate

GUI Odyssey

• Task Type: Mobile GUI navigation
• Output: Action (CLICK/TEXT/SCROLL) + coordinates
• Evaluation: Action accuracy, coordinate error

---

Citation

@dataset{omniagentbench_2026,
  title = {OmniAgentBench: Wild Multimodal Agent Benchmark},
  author = {Hoda Fakharzadeh and Team},
  year = {2026},
  publisher = {HuggingFace Datasets},
  url = {https://huggingface.co/datasets/omniagentbench/OmniAgentBench}
}

---

Contact

For issues or questions, please open an issue on the OmniAgentBench repository.