Voxel51/mind2web_multimodal_test_domain

Dataset Card for "Cross-Domain" Test Split in Multimodal Mind2Web

Note: This dataset is the test split of the Cross-Domain dataset introduced in the paper.

This is a FiftyOne dataset with 4050 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/mind2web_multimodal_test_domain")

# Launch the App
session = fo.launch_app(dataset)

Dataset Description

Curated by: The Ohio State University NLP Group (OSU-NLP-Group)
Shared by: OSU-NLP-Group on Hugging Face
Language(s) (NLP): en
License: OPEN-RAIL License

Dataset Sources

Repository: https://github.com/OSU-NLP-Group/SeeAct and https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web
Paper: "GPT-4V(ision) is a Generalist Web Agent, if Grounded" by Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, Yu Su
Demo: https://osu-nlp-group.github.io/SeeAct

Uses

Direct Use

• Evaluating web agents' ability to generalize to entirely new domains
• Testing zero-shot domain transfer capabilities of models
• Benchmarking the true generalist capabilities of web agents
• Assessing model performance in unseen web environments

Out-of-Scope Use

• Developing web agents for harmful purposes (as stated in the paper's impact statement)
• Automating actions that could violate website terms of service
• Creating agents that access users' personal profiles or perform sensitive operations without consent

Dataset Structure

• Contains 694 tasks across 13 domains and 53 websites
• Tasks average 5.9 actions each
• Average 4,314 visual tokens per task
• Average 494 HTML elements per task
• Average 91,163 HTML tokens per task
• Each example includes task descriptions, HTML structure, operations (CLICK, TYPE, SELECT), target elements with attributes, and action histories

FiftyOne Dataset Structure

Basic Info: 1,338 web UI screenshots with task-based annotations

Core Fields:

• action_uid: StringField - Unique action identifier
• annotation_id: StringField - Annotation identifier
• target_action_index: IntField - Index of target action in sequence
• ground_truth: EmbeddedDocumentField(Detection) - Element to interact with:
  • label: Action type (TYPE, CLICK)
  • bounding_box: a list of relative bounding box coordinates in [0, 1] in the following format: <top-left-x>, <top-left-y>, <width>, <height>]
  • target_action_reprs: String representation of target action
• website: EmbeddedDocumentField(Classification) - Website name
• domain: EmbeddedDocumentField(Classification) - Website domain category
• subdomain: EmbeddedDocumentField(Classification) - Website subdomain category
• task_description: StringField - Natural language description of the task
• full_sequence: ListField(StringField) - Complete sequence of actions for the task
• previous_actions: ListField - Actions already performed in the sequence
• current_action: StringField - Action to be performed
• alternative_candidates: EmbeddedDocumentField(Detections) - Other possible elements

Dataset Creation

Curation Rationale

The Cross-Domain split was specifically designed to evaluate an agent's ability to generalize to entirely new domains it hasn't encountered during training, representing the most challenging generalization scenario.

Source Data

Data Collection and Processing

• Based on the original MIND2WEB dataset
• Each HTML document is aligned with its corresponding webpage screenshot image
• Underwent human verification to confirm element visibility and correct rendering for action prediction
• Specifically includes websites from top-level domains held out from the training data

Who are the source data producers?

Web screenshots and HTML were collected from 53 websites across 13 domains that were not represented in the training data.

Annotations

Annotation process

Each task includes annotated action sequences showing the correct steps to complete the task. These were likely captured through a tool that records user actions on websites.

Who are the annotators?

Researchers from The Ohio State University NLP Group or hired annotators, though specific details aren't provided in the paper.

Personal and Sensitive Information

The dataset focuses on non-login tasks to comply with user agreements and avoid privacy issues.

Bias, Risks, and Limitations

• This split presents the most challenging generalization scenario as it tests performance on entirely unfamiliar domains
• In-context learning methods with large models show better performance than supervised fine-tuning on this split
• The gap between SEEACTOracle and other methods is largest in this split (23.2% step success rate difference)
• Website layouts and functionality may change over time, affecting the validity of the dataset
• Limited to the specific domains included; may not fully represent all possible web domains

Citation

BibTeX:

@article{zheng2024seeact,
  title={GPT-4V(ision) is a Generalist Web Agent, if Grounded},
  author={Boyuan Zheng and Boyu Gou and Jihyung Kil and Huan Sun and Yu Su},
  booktitle={Forty-first International Conference on Machine Learning},
  year={2024},
  url={https://openreview.net/forum?id=piecKJ2DlB},
}

@inproceedings{deng2023mindweb,
  title={Mind2Web: Towards a Generalist Agent for the Web},
  author={Xiang Deng and Yu Gu and Boyuan Zheng and Shijie Chen and Samuel Stevens and Boshi Wang and Huan Sun and Yu Su},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023},
  url={https://openreview.net/forum?id=kiYqbO3wqw}
}

APA:

Zheng, B., Gou, B., Kil, J., Sun, H., & Su, Y. (2024). GPT-4V(ision) is a Generalist Web Agent, if Grounded. arXiv preprint arXiv:2401.01614.

Dataset Card Contact

GitHub: https://github.com/OSU-NLP-Group/SeeAct