PageGuide
Collection
A browser extension that grounds LLM answers in a page directly highlights evidence, navigates step-by-step, hides distractions, and reads Papers • 5 items • Updated • 2
Name stringclasses 6
values | Level stringclasses 2
values | Task stringclasses 7
values | Website URL stringclasses 6
values | Ground Truth float64 | Reference Length float64 |
|---|---|---|---|---|---|
Google Sheet | Medium | How can you highlight duplicate values in a column using conditional formatting? | https://docs.google.com/spreadsheets/u/0/ | null | null |
Google Sheet | Easy | How can you freeze the first row in a Google Sheet so it stays visible when scrolling? | https://docs.google.com/spreadsheets/u/0/ | null | null |
Google Docs | null | How can you add page numbers to a Google Doc but start numbering from page 2 instead of page 1? | https://docs.google.com/document/u/0/ | null | null |
Google Slides | null | How can you change the position of slide numbers for all slides in a Google Slides presentation? | https://docs.google.com/presentation/u/0/ | null | null |
Coda | null | How can you create a view that only shows tasks marked as “Completed”? | https://coda.io/ | null | null |
TradingView | null | How can you move the price scale so it appears on the left side of the chart instead of the default position? | https://www.tradingview.com/chart/ | null | null |
Scratch | null | Make the sprite move to the mouse pointer continuously. | https://scratch.mit.edu/projects/1286152835/editor/ | null | null |
PageGuide Dataset
This repository contains the dataset for PageGuide, a browser extension that assists users in navigating webpages and locating information by grounding LLM answers directly in the HTML DOM.
- Project Page: pageguide.github.io
- Paper: PageGuide: Browser extension to assist users in navigating a webpage and locating information
- Code: github.com/tin-xai/pageguide
Dataset Description
The PageGuide evaluation utilizes several distinct datasets representing different tasks:
pageguide_userstudy: Raw interaction logs from the user study — completion times, chat transcripts, correctness labels, paired statistical results, and post-study survey responses.pageguide_find_data: Task stimuli for the Find condition — 10 real webpages (NASA, Wikipedia, Cleveland Clinic, WWF, Britannica, JMLR) each annotated with up to 2 factual questions, ground-truth answers, and supporting evidence spans.pageguide_guide_data: Task stimuli for the Guide condition — 7 procedural tasks across 6 platforms (Google Sheets, Google Docs, Google Slides, Coda, TradingView, Scratch), labelled Easy or Medium difficulty.pageguide_hide_data: Task stimuli for the Hide condition — 37 annotated webpage snapshots (Amazon, Netflix, TechCrunch, Allrecipes, Spotify, Yelp, and more) with(user_goal, hide_query, difficulty, hidden_elements)annotations and ground-truth CSS selectors.
Sample Usage
You can load these datasets using the Hugging Face datasets library:
User Study Data
from datasets import load_dataset
tasks = load_dataset("ttn0011/pageguide_userstudy", data_files="tasks.csv", split="train").to_pandas()
paired = load_dataset("ttn0011/pageguide_userstudy", data_files="paired_times.csv", split="train").to_pandas()
Find Task Data
from datasets import load_dataset
find_tasks = load_dataset("ttn0011/pageguide_find_data", split="train").to_pandas()
Guide Task Data
from datasets import load_dataset
guide_tasks = load_dataset("ttn0011/pageguide_guide_data", split="train").to_pandas()
Hide Task Data
from datasets import load_dataset
hide_tasks = load_dataset("ttn0011/pageguide_hide_data", split="train").to_pandas()
- Downloads last month
- 23