Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
10
10
url
stringclasses
33 values
goal
stringlengths
12
949
source
stringclasses
3 values
split
stringclasses
1 value
task-00000
https://dictionary.cambridge.org/
Learn how to use the dictionary translation.
nnetnav
train
task-00001
https://www.amazon.com/
Search for a laptop on Amazon.
nnetnav
train
task-00002
https://arxiv.org/
Find the abstract of the "Pilot-Quantum A Quantum-HPC Middleware for Resource, Workload and Task Management" research paper on the Computing Research Repository.
nnetnav
train
task-00003
https://www.google.com/travel/flights
Search for flights or set travel dates for a trip.
nnetnav
train
task-00004
https://www.apple.com/
Find the price of iPhone 16.
nnetnav
train
task-00005
https://www.coursera.org/
Find the "Arts and Culture Strategy" course on Coursera.
nnetnav
train
task-00006
https://github.com/
Investigate recent issues with GitHub's systems.
nnetnav
train
task-00007
https://www.coursera.org/
Find beginner-level courses on teaching methods.
nnetnav
train
task-00008
https://www.booking.com/
Find available hotels in Las Vegas from 1 January to 3 January 2025
nnetnav
train
task-00009
https://dictionary.cambridge.org/
Find out the meaning of the phrase "in a nutshell".
nnetnav
train
task-00010
https://www.allrecipes.com/
Find different recipe ideas for baked goods and meals.
nnetnav
train
task-00011
https://arxiv.org/
Get the details on the data acquisition process in the research paper "AI-Powered Intracranial Hemorrhage Detection A Co-Scale Convolutional Attention Model with Uncertainty-Based Fuzzy Integral Operator and Feature Screening."
nnetnav
train
task-00012
https://www.coursera.org/
Find Coursera's professional certificates.
nnetnav
train
task-00013
https://huggingface.co/
Search for the details about the "HuggingFaceTB/finemath" dataset on Hugging Face.
nnetnav
train
task-00014
https://github.com/
Find the cost of using GitHub Codespaces.
nnetnav
train
task-00015
https://www.coursera.org/
Find a beginner-friendly course on Coursera for Mathematics that is related to the topic of Introduction to Advanced Calculus.
nnetnav
train
task-00016
https://www.google.com/travel/flights
Search for flights from New York City to Las Vegas on January 1-3, 2025.
nnetnav
train
task-00017
https://www.google.com/
Research The Terai region in Nepal for travel purposes, perhaps to plan a trip.
nnetnav
train
task-00018
https://www.espn.com/
Check NFL game scores after viewing college football scores.
nnetnav
train
task-00019
https://www.wolframalpha.com/
Find temperature data related to climate models.
nnetnav
train
task-00020
https://github.com/
Find a job opening at GitHub.
nnetnav
train
task-00021
https://www.bbc.com/news
Research the impact of social issues or trends (e.g. loneliness, relationships between parents and teachers) as a potential inspiration for writing.
nnetnav
train
task-00022
https://www.booking.com/
Plan a trip to Las Vegas.
nnetnav
train
task-00023
https://huggingface.co/
Get the license for answerdotai/ModernBERT-base
nnetnav
train
task-00024
https://www.espn.com/
Get scores of multiple games involving USC.
nnetnav
train
task-00025
https://www.coursera.org/
Find a course on Financial Markets.
nnetnav
train
task-00026
https://huggingface.co/models
Find a model that can perform text generation or sentiment analysis tasks.
nnetnav
train
task-00027
https://arxiv.org/
Find the research on "Statistics of Turbulence from Spectral-Line Data Cubes" on arXiv.org.
nnetnav
train
task-00028
https://github.com/
Learn about DevOps on GitHub.
nnetnav
train
task-00029
https://www.google.com/
Find out about global warming and its effects on the environment.
nnetnav
train
task-00030
https://huggingface.co/
Find a Text-to-Image model.
nnetnav
train
task-00031
https://www.coursera.org/
Find information about a course on leadership.
nnetnav
train
task-00032
https://www.allrecipes.com/
Find the recipe for "Juicy Roasted Chicken".
nnetnav
train
task-00033
https://www.google.com/travel/flights
Set travel dates to January 1, 2025 and January 3, 2025.
nnetnav
train
task-00034
https://www.apple.com/
Research the Apple Watch Ultra 2.
nnetnav
train
task-00035
https://www.bbc.com/news
Find out more about wind farms and climate change efforts in the UK.
nnetnav
train
task-00036
https://arxiv.org/
Find videos on simplified physics experiments.
nnetnav
train
task-00037
https://arxiv.org/
Find a recent research paper on machine learning.
nnetnav
train
task-00038
https://www.apple.com/
Check the prices and features of different iPhone models and AirPods on Apple's website.
nnetnav
train
task-00039
https://www.wolframalpha.com/
Find information about the chemical properties and structure of water using WolframAlpha.
nnetnav
train
task-00040
https://www.amazon.com/
Find the cheapest hand sanitizer.
nnetnav
train
task-00041
https://www.coursera.org/
Explore Coursera's Data Science courses and career paths.
nnetnav
train
task-00042
https://www.google.com/
Find more information about the Cybersecurity major at Michigan Technological University.
nnetnav
train
task-00043
https://www.espn.com/
Find the injury report for the Falcons vs Commanders game.
nnetnav
train
task-00044
https://www.apple.com/
Find out the price of HomePod (2nd generation).
nnetnav
train
task-00045
https://arxiv.org/
Find the related work discussed in a research article on Computer Vision.
nnetnav
train
task-00046
https://www.booking.com/
Find a hotel room in Las Vegas from January 1-8, 2025.
nnetnav
train
task-00047
https://www.apple.com/
Check if an Apple device is still under warranty.
nnetnav
train
task-00048
https://www.allrecipes.com/
Find a kid-friendly, Paleo banana pancake recipe to make using only a few ingredients.
nnetnav
train
task-00049
https://huggingface.co/
Search for information related to business syllabus using datasets available on Hugging Face.
nnetnav
train
task-00050
https://www.allrecipes.com/
Find Christmas recipes.
nnetnav
train
task-00051
https://dictionary.cambridge.org/
Learn the difference between the US and UK pronunciations of the word "hello".
nnetnav
train
task-00052
https://www.apple.com/
Compare the features of the Apple Watch Series 10 with other models.
nnetnav
train
task-00053
https://www.wolframalpha.com/
Find the date and time of the next full moon in 2026.
nnetnav
train
task-00054
https://www.coursera.org/
Find a project management course by Google.
nnetnav
train
task-00055
https://www.wolframalpha.com/
Learn about the Riemann Hypothesis.
nnetnav
train
task-00056
https://dictionary.cambridge.org/
What are some collocations of accommodation, specifically affordable accommodation?
nnetnav
train
task-00057
https://www.espn.com/
Explore ESPN for sports updates and news, primarily in college football.
nnetnav
train
task-00058
https://www.google.com/
Find publications on responsible AI.
nnetnav
train
task-00059
https://www.google.com/maps
Get the driving distance from Oakland to San Francisco.
nnetnav
train
task-00060
https://www.espn.com/
Find the main sections of the ESPN website
nnetnav
train
task-00061
https://www.google.com/
Find out what foods are good for high blood pressure (or possibly find out what foods can help to manage the condition specifically with regards to a specific ingredient that may help such as carrots).
nnetnav
train
task-00062
https://www.amazon.com/
Choose a gift card design for a graduation.
nnetnav
train
task-00063
https://www.google.com/maps
Find a Japanese restaurant in Buenos Aires.
nnetnav
train
task-00064
https://www.bbc.com/news
Find the latest news about the Israel and Gaza conflict.
nnetnav
train
task-00065
https://huggingface.co/
Find the Facebook mbart large 50 many to many mmt model.
nnetnav
train
task-00066
https://www.bbc.com/news
Find out the current stock market news from the BBC website.
nnetnav
train
task-00067
https://dictionary.cambridge.org/
Investigate the meaning of feeling sad and the synonyms of the word depression.
nnetnav
train
task-00068
https://github.com/
Find out how to upgrade from the free version of GitHub Copilot to the pro version.
nnetnav
train
task-00069
https://huggingface.co/
Find the latest version of the Hugging Face Hub client library.
nnetnav
train
task-00070
https://www.wolframalpha.com/
Find the expansion of the polynomial (x^2 + 1)(x^2 - 1)(x+1)^3
nnetnav
train
task-00071
https://arxiv.org/
Explore topics related to the simplified quantum physics lecture for high school on arXiv.
nnetnav
train
task-00072
https://www.amazon.com/
Find the best strength training bench on Amazon.
nnetnav
train
task-00073
https://www.wolframalpha.com/
Find data on the salary of a data scientist in the United States, including historical data.
nnetnav
train
task-00074
https://www.google.com/maps
Find moderately priced bars in New York City with at least 4.1 stars rating.
nnetnav
train
task-00075
https://github.com/
Find information about data structures within the Python implementation in TheAlgorithms repository.
nnetnav
train
task-00076
https://dictionary.cambridge.org/
Find the translation of the word "jukebox" in Simplified Chinese.
nnetnav
train
task-00077
https://www.amazon.com/
Find a budget gift idea for her on Amazon.
nnetnav
train
task-00078
https://www.google.com/maps
Find the schedule and directions for taking public transportation from San Francisco to Palo Alto.
nnetnav
train
task-00079
https://www.google.com/travel/flights
Find the cheapest flight from San Francisco to Tampa.
nnetnav
train
task-00080
https://www.booking.com/
Search for hotels in New York for January 2025.
nnetnav
train
task-00081
https://huggingface.co/
Find the base models of the language model that incorporates company-related factual knowledge, created by "sophia-jihye".
nnetnav
train
task-00082
https://www.espn.com/
Find the current AFC East standings in the 2024 NFL season.
nnetnav
train
task-00083
https://www.google.com/
View Google Trends for different regions.
nnetnav
train
task-00084
https://www.google.com/
Get resources on parenting advice, child development, and activities for a 2-year-old.
nnetnav
train
task-00085
https://huggingface.co/
Troubleshoot the "Task not found for this model" error in a Hugging Face model after completing the NLP course.
nnetnav
train
task-00086
https://www.coursera.org/
Compare the features and requirements of different project management courses and degree programs on Coursera.
nnetnav
train
task-00087
https://www.google.com/maps
Find the best hotel deals from San Francisco to Palo Alto.
nnetnav
train
task-00088
https://www.wolframalpha.com/
Get a Wolfram format of the plot of the derivative of tan(x).
nnetnav
train
task-00089
https://www.google.com/maps
Find moderately priced French restaurants in the Williamsburg area of New York City.
nnetnav
train
task-00090
https://www.allrecipes.com/
Find kid-friendly snack ideas for Halloween.
nnetnav
train
task-00091
https://www.allrecipes.com/
Find different recipes, including BBQ sauce and grilled vegetables, to plan a meal.
nnetnav
train
task-00092
https://www.bbc.com/news
Find recent business news about Boeing.
nnetnav
train
task-00093
https://www.amazon.com/
Look for cookbooks with dinner recipes on Amazon.
nnetnav
train
task-00094
https://www.coursera.org/
Find a beginners python course related to data science that has a business focus
nnetnav
train
task-00095
https://www.bbc.com/news
What are some popular travel destinations in India that inspired M T Vasudevan Nair's writing?
nnetnav
train
task-00096
https://www.espn.com/
Get the Buffalo Bills' NFL standings for Week 18
nnetnav
train
task-00097
https://dictionary.cambridge.org/
Find information about adverb phrases.
nnetnav
train
task-00098
https://www.bbc.com/news
Read recent news articles about the Middle East.
nnetnav
train
task-00099
https://dictionary.cambridge.org/
Find the definitions of multiple vocabulary words using the Cambridge Dictionary website.
nnetnav
train
End of preview. Expand in Data Studio

wev data

Jun Huang*, Xin Ren* · University of Electronic Science and Technology of China · *Equal contribution

This dataset holds the browser-step data behind the wev decision models: wev-1.7b, wev-4b and wev-8b. Every row is one browser step, written as a POST /v1/systemone request (a state plus typed questions) with its labelled answers. The requests use exactly the format the open browser agent jev-ultrafast sends to its System One, so a decision model trained here can be served behind that agent unchanged.

Subset Rows (train / validation / test) Source License
mind2web 5,863 / 586 / 875 Mind2Web, converted CC BY 4.0
nnetnav_audited 11,408 / 1,140 / 1,150 NNetNav-live, converted, stopping labels audited Apache-2.0
teacher_first 2,548 / 199 / – Episodes of a prompted LLM (qwen3-max) acting as System One on live websites see Terms
teacher_second 2,274 / 248 / – A second collection on the tasks the first had not solved see Terms
live_tasks 1,402 / – / 153 Goals and start URLs for live-website runs; test is the held-out end-to-end suite see Terms

The general typed-decision corpora used in training (Kev decision-v7, typed-decisions, tasksource-jev, jev-distill-corpus-v3, typed-decisions-synth) are not redistributed here; the repository's builders download and convert them from their sources.

Format

{
  "request": {
    "model": "wev-latest",
    "state": {
      "page": {"url": "...", "title": "...", "text": "visible page text"},
      "elements": [{"index": "1", "role": "button", "label": "Search", "operations": ["CLICK"]}],
      "recent_actions": [{"action": "...", "kind": "click", "text": null, "page_changed": true}]
    },
    "questions": {
      "operation": {"type": "choice",
                    "instructions": {"goal": "...", "rules": ["..."]},
                    "criteria": {"CLICK": "...", "TYPE_TEXT": "...", "DONE": "...", "BLOCKED": "..."}},
      "click_target": {"type": "choice",
                       "instructions": {"goal": "...", "operation": "CLICK", "rules": ["..."]},
                       "criteria": {"1": {"element": "[1] Search", "current_value": "", "role": "button"}}}
    }
  },
  "labels": {"operation": "CLICK", "click_target": "1"},
  "_meta": {"source": "..."}
}

A step asks for the operation and, for each operation that needs one, a target (click_target, type_text_target or select_target). labels holds the reference option key for each labelled question. A step is answered correctly when the operation and, if the operation takes one, its target both match. Targets are chosen among the 8–40 candidate elements listed in the state, not among every element on the page. In teacher rows, _meta records the episode, the task and the teacher's one-sentence reason for its choice.

from datasets import load_dataset
steps = load_dataset("alanhuangya/wev-data", "nnetnav_audited", split="test")

To train with the wev package, download the files and pass the subset folders to wev train --data; validation.jsonl is read as the development split.

How the subsets were built

mind2web. Each recorded step becomes one request containing the page's visible text, a sample of 8–40 candidate elements that includes the gold element, the task and the preceding actions. Splits are by website, so every test website is unseen in training.

nnetnav_audited. NNetNav supplies the stopping (DONE), giving-up (BLOCKED) and scrolling steps that Mind2Web lacks. Its trajectories come from unsupervised exploration, so its stopping labels are noisy. An LLM judge (DeepSeek-V4.1-Flash) read the goal and the state at every step and decided whether the goal was already achieved. On the training split it rejected 592 of 1,898 DONE labels (31%) and found the goal already achieved at 2,281 of the 10,102 other steps (23%). Those steps were relabelled DONE and the rejected DONE steps were dropped. The validation and test splits were audited the same way.

teacher_first, teacher_second. A prompted LLM (qwen3-max) served as the System One of jev-ultrafast on live websites. Its prompt forbade signing in, registering, buying, booking, posting, messaging and submitting personal information, and asked it to answer BLOCKED on CAPTCHAs, unusual-traffic pages and login walls. Every request and the teacher's choice were logged. An LLM judge then read the final page of each episode, and the kept rows are:

  • every decision of an episode whose DONE the judge confirmed;
  • every decision up to a BLOCKED on a page that blocked the agent;
  • every decision but the last of an episode cut short by the browser harness.

Episodes that looped or exhausted their budget were dropped, as were repeated states. The judge rejected 155 of the teacher's 431 DONE claims (36%). Train and validation are split by task.

live_tasks. Goals paired with start URLs: NNetNav goals on their live sites, Wikipedia look-ups and Google Flights searches. The 153 test tasks are the end-to-end suite used to evaluate the wev models; none of them appears in the teacher subsets.

Terms

The Mind2Web-derived subset follows CC BY 4.0 and the NNetNav-derived subset follows Apache-2.0. Attribute the original datasets (see Citation). The teacher subsets and the live tasks contain text captured from public websites, which remains subject to those sites' terms, and the teacher's choices are outputs of qwen3-max, which are subject to its provider's terms. Treat these subsets as research data, and check the applicable terms before any commercial use. The data describes browser steps only and contains no credentials; the teacher was instructed never to sign in to any site.

Citation

Jun Huang and Xin Ren contributed equally (University of Electronic Science and Technology of China).

@misc{huang2026wev,
  title     = {wev: Distilling LLM Browser Agents into Open, Local System-One Decision Models},
  author    = {Huang, Jun and Ren, Xin},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22941164},
  url       = {https://doi.org/10.5281/zenodo.22941164}
}

If you use the converted subsets, please also cite their sources:

@inproceedings{deng2023mind2web,
  title     = {Mind2Web: Towards a Generalist Agent for the Web},
  author    = {Xiang Deng and Yu Gu and Boyuan Zheng and Shijie Chen and Samuel Stevens and Boshi Wang and Huan Sun and Yu Su},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2023}
}
@misc{murty2024nnetnav,
  title         = {NNetNav: Unsupervised Learning of Browser Agents Through Environment Interaction in the Wild},
  author        = {Shikhar Murty and Hao Zhu and Dzmitry Bahdanau and Christopher D. Manning},
  year          = {2024},
  eprint        = {2410.02907},
  archivePrefix = {arXiv}
}
Downloads last month
32

Space using alanhuangya/wev-data 1

Collection including alanhuangya/wev-data

Paper for alanhuangya/wev-data