MC-Mind2Web

MC-Mind2Web is constructed from Mind2Web(Multimodal-Mind2Web) using a semi-automated, action-centric, three-step annotation pipeline designed to facilitate research on our introduced new task: Multimodal Proactive Web Navigation.

Statistics

Statistics	Train	Test (Task)	Test (Website)	Test (Domain)
Task Distribution
# Tasks	1,009	177	142	694
# Websites	73	64	10	53
# Subdomains	17	17	9	13
Avg # Actions	7.71	7.56	7.18	5.85
Value Identification
Avg # Values	3.04	3.02	3.19	2.38
Max # Values	11	9	12	10
Avg # Selective	2.37	2.30	2.50	1.90
Avg # Multi-Turn	1.61	1.63	1.58	1.20
Task Complexity
# Complete	100	17	14	69
# Easy	302	53	42	208
# Medium	302	53	42	208
# Hard	305	54	44	209
Value Modification
# Omission	629	114	83	423
# Generalization	606	102	90	436
# Unfeasible	586	105	85	392
# Total	1,821	321	258	1,251
Conversation Characteristics
Avg # Conversation Turns	1.80	1.81	1.82	1.80
Avg Question Length	28.39	28.24	28.49	28.47
Avg Answer Length	7.85	7.60	7.62	7.78

Fields

Identifiers:

• action_uid (str): unique id for each action (step)
• annotation_id (str): unique id for each task

Context:

• domain (str): website domain
• subdomain (str): website subdomain
• website (str): website name

Visual:

• screenshot (image): webpage screenshot image corresponding to the HTML (loaded as PIL Image object)

HTML Content:

• pos_candidates (list[str]): ground truth elements. Here we only include positive elements that exist in "cleaned_html" after our preprocessing, so "pos_candidates" might be empty. The original labeled element can always be found in the "raw_html".
  • tag (str): tag of the element
  • is_original_target (bool): whether the element is the original target labeled by the annotator
  • is_top_level_target (bool): whether the element is a top level target find by our algorithm. please see the paper for more details
  • backend_node_id (str): unique id for the element
  • attributes (str): serialized attributes of the element, use json.loads to convert back to dict
• neg_candidates (list[str]): other candidate elements in the page after preprocessing, has similar structure as "pos_candidates"
• cleaned_html (str): cleaned html of the page before the action is performed
• raw_html (str): raw html of the page before the action is performed

Task Information (Original):

• confirmed_task (str): task description
• action_reprs (list[str]): human readable string representation of the action sequence
• target_action_index (str): the index of the target action in the action sequence
• target_action_reprs (str): human readable string representation of the target action
• operation (str): operation to perform:
  • op (str): operation type, one of CLICK, TYPE, SELECT
  • original_op (str): original operation type, contain additional HOVER and ENTER that are mapped to CLICK, not used
  • value (str): optional value for the operation, e.g., text to type, option to select

Task Information (MC-Mind2Web):

• vague_task (str): Vague version of the task instruction
• difficulty (str): Task difficulty level ("easy", "medium", "hard")

Action Information (MC-Mind2Web):

• dependent_value (str or None): Value that the action depends on
• selective_action (bool or None): Whether action requires selection from multiple options
• clarification_point (bool): Whether this action requires clarification
• clarification_details (str or None): Clarification dialogue details
  • modification_type (str): Type of modification ("omission", "generalization", or "unfeasible")
  • modified_value (str): value after modification
  • clarifying_question (str): Question asked to the user for clarification
  • user_response (str): User's response to the clarifying question

Visual Information (MC-Mind2Web):

• crop_bbox (str or None): Cropping information for visual elements
  • top (int): Top boundary of the screenshot crop
  • bottom (int): Bottom boundary of the screenshot crop
  • image_width (int): Width of the screenshot
  • image_height (int): Height of the screenshot
  • type (str): Type of bbox ("no_bbox", "invalid_bbox", or "valid_bbox")
  • element_bbox (dict or None): Bounding box of the target element
    • x (float): X coordinate of the element
    • y (float): Y coordinate of the element
    • width (float): Width of the element
    • height (float): Height of the element

Usage

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("ymwangv/MC-Mind2Web")

# Load specific split
train_dataset = load_dataset("ymwangv/MC-Mind2Web", split="train")