VerdictBench

VerdictBench is the first large-scale multimodal benchmark for e-commerce dispute verdicts (EDV), introduced in CyberJurors: A Multi-Agent Simulation Task for E-Commerce Disputes Verdict (ICML 2026).

It contains 6,000 real-world cases collected from public e-commerce dispute logs. Each case includes:

• transaction metadata
• multi-round textual claims from buyer and seller
• multimodal evidence (chat text, images, videos)
• the ground-truth verdict of 17 crowdsourced human jurors

Cases span 5 top-level categories and are split (3:1:2) by category + vote-margin difficulty into train / val / test.

Access

This repository is gated. Click “Request access”, fill in your affiliation and intended use, and wait for manual approval. After approval, log in with your HF token to download:

huggingface-cli login
from datasets import load_dataset
ds = load_dataset("piggi/VerdictBench")

Intended use

Academic research on multimodal evidence reasoning, jury simulation, and fair automated dispute verdicts. Not for commercial deployment or re-identification of disputants.

Citation

@article{sun2026cyberjurors,
  title={CyberJurors: A Multi-Agent Simulation Task for E-Commerce Disputes Verdict},
  author={Sun, Yanhui and Liu, Wu and Ming, Haifeng and Wang, Xinru and Yao, Hantao and Zhang, Yongdong},
  journal={arXiv preprint arXiv:2605.28369},
  year={2026}
}