metadata
dataset_info:
- config_name: action
features:
- name: inputs
dtype: string
- name: labels
dtype: string
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 544845750
num_examples: 6108
- name: validation
num_bytes: 64700486
num_examples: 773
- name: test
num_bytes: 86931891
num_examples: 894
download_size: 172080282
dataset_size: 696478127
- config_name: dom-element
features:
- name: inputs
dtype: string
- name: labels
dtype: int64
- name: metadata
dtype: string
- name: choices_list
list: string
splits:
- name: train
num_bytes: 370787325
num_examples: 4040
- name: validation
num_bytes: 41059613
num_examples: 516
- name: test
num_bytes: 51300842
num_examples: 553
download_size: 116081182
dataset_size: 463147780
configs:
- config_name: action
data_files:
- split: train
path: action/train-*
- split: validation
path: action/validation-*
- split: test
path: action/test-*
- config_name: dom-element
data_files:
- split: train
path: dom-element/train-*
- split: validation
path: dom-element/validation-*
- split: test
path: dom-element/test-*
license: cc-by-4.0
source_datasets:
- osunlp/Mind2Web
mind2web-dom
Mind2Web public training trajectories as compact HTML-only actions and native MCQs.
Dataset annotations are CC BY 4.0: https://huggingface.co/datasets/osunlp/Mind2Web No official benchmark test files are downloaded or republished. Derived 80/10/10 train/validation/test partitions group whole trajectories and identical cleaned DOM snapshots. They are internal holdouts, not the official Mind2Web benchmark splits. Only past actions enter context. Four-way MC uses the original target and three deterministically sampled native negatives, excluding all alternative positives.
Original data: osunlp/Mind2Web. Repackaged as parquet for tasksource by scripts/repackage_dataset/.