NEU-HAI/OPeRA

Dataset Card for Dataset Name

Dataset Summary

OPeRA is a dataset of Observation, Persona, Rationale, and Action collected from real human users during online shopping. OPeRA provides rich, time-aligned logs of users' web browsing behavior, completed with self-reported rationales and detailed self-reported persona profiles. Unlike prior datasets, OPeRA captures not only what users do but also why they do it, enabling deeper insights into decision-making processes. To ensure user privacy and data safety, the released dataset has undergone rigorous filtering and de-identification. It contains 692 shopping sessions from 51 unique users, 28,904 <observation, action> pairs, and 604 human-annotated rationales. OPeRA serves as the first benchmark dataset for evaluating LLM agents on personalized and verifiable user behavior simulation.

Dataset Versions

This dataset comes in two versions:

1. Full Version : Includes the complete action space with all captured user interactions: 'click', 'scroll', 'input', 'navigation', and 'tabActivate'.

2. Filtered Version : Applies a simplification of the action space, leaving action types including 'click', 'input', and 'terminate'.

(The corresponding screenshots will be released soon)

Dataset Structure

Both versions are organized into three main components stored as Parquet files:

OPeRA
├── full_user
│   ├── train
│   └── test
├── full_session
│   ├── train
│   └── test
├── full_action
│   ├── train
│   └── test
├── filtered_user
│   ├── train
│   └── test
├── filtered_session
│   ├── train
│   └── test
└── filtered_action
    ├── train
    └── test    

Data Fields

User Table

Contains user-level information including surveys and interviews:

• user_id (str): Anonymized unique identifier for each user
• survey (str): JSON string containing organized user survey responses about shopping preferences, demographics, and personality
• raw_survey (str): JSON string containing raw user survey responses
• interview_transcript (str): Transcript of user interview
• interview_transcript_processed (str): LLM-summarized version of the interview transcript

Session Table

Contains session-level information:

• session_id (str): Unique identifier for each shopping session (format: {user_id}_{start_time}_{end_time})
• user_id (str): References the user who performed this session
• action_count (int): Number of actions in this session

Action Table - Full Version

Contains action-level information with detailed interaction data for the full version:

Fields:

• session_id (str): References the session this action belongs to
• action_id (str): Unique identifier for each action
• timestamp (str): Action timestamp
• action_type (str): Type of action: click, scroll, input, navigation, tab_activate
• url (str): Masked web URL where the action occurred
• window_size (str): JSON string containing user's browser window dimensions
• page_meta (str): JSON string containing page metadata with masked URLs
• simplified_html (str): Simplified version of HTML content before the action happens
• rationale (str): User-provided rationale for this action (if applicable)
• products (str): JSON string containing information about list of products involved in this action (if applicable). Each product has fields:
  • asin (str): Amazon Standard Identification Number
  • title (str): Product title
  • price (str): Product price
  • quantity (int): Purchase quantity
  • options (list): Product options

Action-Type Specific Fields:

For action_type == "click":

• click_type (str): Classification of click type. Possible values include:
  • purchase: Click on purchase intention related buttons (checkout, buy now, set subscription, add to cart)
  • search: Click on search buttons or search boxes
  • review: Click on review-related elements
  • filter: Click on filters
  • quantity: Click on quantity-related elements (quantity increase/decrease, delete item)
  • product_option: Click on product option selections (size, color, etc.)
  • cart_side_bar: Click on shopping cart sidebar elements
  • suggested_term: Click on suggested search terms
  • nav_bar: Click on navigation bar elements
  • page_related: Click on pagination elements or carousel navigation buttons
  • cart_page_select: Click on cart page selection elements (e.g. item checkbox)
  • product_link: Click on product links or product images
  • other: Other types of clicks not covered by the above categories
• semantic_id (str): Semantic identifier for the clicked element in simplified HTML
• mouse_position (str): JSON string containing mouse coordinates when clicking
• element_meta (str): JSON string containing element metadata with name and data fields

For action_type == "input":

• semantic_id (str): Semantic identifier for the input element in simplified HTML
• input_text (str): Value entered in the input field
• mouse_position (str): JSON string containing mouse coordinates

For action_type == "scroll":

• scroll_detail (str): JSON string containing scroll information with fields:
  • scroll_current_top (str): Current scroll position from top
  • scroll_current_left (str): Current scroll position from left
  • scroll_distance_x (str): Horizontal scroll distance
  • scroll_distance_y (str): Vertical scroll distance

For action_type == "navigation":

• navigation_type (str): Type of navigation (back, forward, reload, new page)
• simplified_html_new_page (str): Simplified HTML content of the new page after navigation

For action_type == "tab_activate":

• simplified_html_new_page (str): Simplified HTML content of the new page after navigation

Action Table - Filtered Version

Contains action-level information with detailed interaction data for the filtered version:

Fields:

• session_id (str): References the session this action belongs to
• action_id (str): Unique identifier for each action
• timestamp (str): Action timestamp
• action_type (str): Type of action: click, input, terminate
• url (str): Masked web URL where the action occurred
• window_size (str): JSON string containing user's browser window dimensions
• page_meta (str): JSON string containing page metadata with masked URLs
• simplified_html (str): Simplified version of HTML content for easier parsing
• rationale (str): User-provided rationale for this action (if applicable)
• products (str): JSON string containing information about list of products involved in this action (if applicable). Each product has fields:
  • asin (str): Amazon Standard Identification Number
  • title (str): Product title
  • price (str): Product price
  • quantity (int): Purchase quantity
  • options (list): Product options

Action-Type Specific Fields:

For action_type == "click":

• click_type (str): Classification of click type. Possible values include:
  • purchase: Click on purchase intention related buttons (checkout, buy now, set subscription, add to cart)
  • search: Click on search buttons or search boxes
  • review: Click on review-related elements
  • filter: Click on filters
  • quantity: Click on quantity-related elements (quantity increase/decrease, delete item)
  • product_option: Click on product option selections (size, color, etc.)
  • cart_side_bar: Click on shopping cart sidebar elements
  • suggested_term: Click on suggested search terms
  • nav_bar: Click on navigation bar elements
  • page_related: Click on pagination elements or carousel navigation buttons
  • cart_page_select: Click on cart page selection elements (e.g. item checkbox)
  • product_link: Click on product links or product images
  • other: Other types of clicks not covered by the above categories
• semantic_id (str): Semantic identifier for the clicked element in simplified HTML
• mouse_position (str): JSON string containing mouse coordinates when clicking
• element_meta (str): JSON string containing element metadata with name and data fields

For action_type == "input":

• semantic_id (str): Semantic identifier for the input element in simplified HTML
• input_text (str): Value entered in the input field
• mouse_position (str): JSON string containing mouse coordinates

For action_type == "terminate":

• No additional fields (indicates end of user session)

Data Distribution

Disclaimer

This dataset is intended for research purposes only. We hope it can contribute to the advancement of the NLP community. Please use it responsibly and refrain from any misuse or applications that could cause harm. By accessing the dataset, you agree to uphold these principles.

Citation Information

@misc{wang2025operadatasetobservationpersona,
      title={OPeRA: A Dataset of Observation, Persona, Rationale, and Action for Evaluating LLMs on Human Online Shopping Behavior Simulation},
      author={Ziyi Wang and Yuxuan Lu and Wenbo Li and Amirali Amini and Bo Sun and Yakov Bart and Weimin Lyu and Jiri Gesi and Tian Wang and Jing Huang and Yu Su and Upol Ehsan and Malihe Alikhani and Toby Jia-Jun Li and Lydia Chilton and Dakuo Wang},
      year={2025},
      eprint={2506.05606},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.05606},
}