Wonderbread (Text Only)

zstd-compressed Parquet conversion of the JSON/Text traces from the WONDERBREAD dataset (Paper, Original Zenodo).

The original 7.4GB of text-only traces has been packed into a single Parquet file; The highly repetitive AXTrees and HTML DOM structures in the dataset compress extremely efficiently.

Data subsets:

• default (viewer-safe): excludes the heavy trace JSON strings to prevent browser OOM on the Hugging Face dataset viewer.
• full: includes all columns, including the raw DOM snapshots and action history in the trace column.

Dataset Structure

Each row represents a single workflow demonstration:

• workflow_id (string): The original trace ID including the timestamp (e.g., 102 @ 2023-12-27-00-57-01)
• task_id (int32): The parsed integer WebArena task ID (e.g., 102)
• intent (string): The natural language intent of the task (e.g., What are the top-2 best-selling product in 2022)
• site (string): The WebArena site (e.g., shopping_admin)
• start_url (string): The starting URL of the workflow
• sop (string): The raw text of the Standard Operating Procedure (SOP - ...txt)
• trace (string, only in full config): The raw stringified JSON array containing the DOM states and actions.

Usage

from datasets import load_dataset
import json

# Lightweight subset (default, viewer-safe)
ds = load_dataset("tkukurin/wonderbread")
train_split = ds["train"]

# Full dataset with the heavy JSON trace strings
ds_full = load_dataset("tkukurin/wonderbread", "full")

# Example: Parsing the JSON trace back into a list of steps
sample_trace_str = ds_full["train"][0]["trace"]
trace_list = json.loads(sample_trace_str)

Citation

@inproceedings{wornow2024wonderbread,
  title={WONDERBREAD: A Benchmark for Evaluating Multimodal Foundation Models on Business Process Management Tasks},
  author={Wornow, Michael and others},
  booktitle={arXiv preprint arXiv:2406.13264},
  year={2024}
}