OmniAgentBench

Benchmarking Multimodal Agent Robustness Under Wild Conditions

A multi-input benchmark where each sample includes text, images, and speech audio — same schema and ground truth for every row.

Dataset Overview

2,700 MPCC samples — one row per (task × difficulty × instance). Each row has full text, spoken text, gold answer, image file names, and audio file name.

Field	Description
sample_id	Unique identifier (e.g. mpcc_flight_easy_1)
task	flight / calendar / meeting
difficulty	easy / medium / hard
category	Task subtype (e.g. flight planning)
text_instruction	Full textual prompt (includes output format instructions)
spoken_text	Natural-language question only (what the WAV says)
gold_answer	Ground-truth answer (JSON string with best and feasible)
image_files	List of JPEG filenames under images/mpcc/
images_available	true when images are part of the release
audio_file	WAV filename under mpcc/
audio_available	true when TTS is part of the release

Samples per split (uniform)

Task	Easy	Medium	Hard	Subtotal
Flight	300	300	300	900
Calendar	300	300	300	900
Meeting	300	300	300	900
Total	900	900	900	2,700

Tabular index: data/dataset_multimodal.parquet and dataset/mpcc/dataset_multimodal.json (same 2,700 rows).

Speech assets: mpcc/*.wav + mpcc/manifest.json (metadata mirrors the JSON rows; manifest is optional if you load from the JSON only).

Audio is synthesized with Qwen3-TTS from spoken_text.

Built on MPCC.

Files

omniagentbench/OmniAgentBench/
├── data/
│   └── dataset_multimodal.parquet   # 2,700 rows
├── dataset/mpcc/
│   ├── dataset_multimodal.json      # 2,700 rows (same as parquet)
│   └── dataset_multimodal_700_backup.json   # legacy 700-row export (optional)
├── images/mpcc/
│   └── *.jpg
└── mpcc/
    ├── manifest.json                # 2,700 entries (aligned with tabular rows)
    └── *.wav                        # 2,700 TTS files

Quick Start

from huggingface_hub import hf_hub_download
import json

REPO = "omniagentbench/OmniAgentBench"

path = hf_hub_download(repo_id=REPO, filename="dataset/mpcc/dataset_multimodal.json", repo_type="dataset")
samples = json.load(open(path))
assert len(samples) == 2700

s = samples[0]
audio = hf_hub_download(repo_id=REPO, filename=f"mpcc/{s['audio_file']}", repo_type="dataset")
for img in s["image_files"]:
    hf_hub_download(repo_id=REPO, filename=f"images/mpcc/{img}", repo_type="dataset")

Or with the datasets library:

from datasets import load_dataset
ds = load_dataset("omniagentbench/OmniAgentBench", split="test")
print(len(ds))   # 2700

Modalities

• Text — full instruction + spoken_text for speech experiments.
• Images — two screenshots per sample (image_files).
• Audio — TTS WAVs in mpcc/, filenames in audio_file.

Citation

@misc{omniagentbench2026,
  title={OmniAgentBench: Measuring Multimodal Agent Robustness Under Wild Conditions},
  author={Fakhar, Hoda and others},
  year={2026},
  url={https://github.com/hodfa840/OmniAgentBench}
}

Links

• Code: github.com/hodfa840/OmniAgentBench
• Organization: huggingface.co/omniagentbench
• Original MPCC: huggingface.co/datasets/jyyyyy67/MPCC