convexworks/cookie-banners · Datasets at Hugging Face

·Dataset Viewer Auto-converted to Parquet Duplicate

·Subset (3) banners · 1.01k rows banners (1.01k rows)model_judgments (1.01k rows)sites (2.36k rows) Split (1) train · 1.01k rows train (1.01k rows)

·Cookie Banner Text and Controls

·The wording, controls and markup of 1,010 cookie banners, taken from the saved homepages of European B2B software companies in June 2026. Each row is one banner: its title, its text, every button, link and toggle on its first screen, and the banner's HTML, in the original wording.

·Use it to study how consent is worded, or to prototype tools that read banner text and button labels. It does not measure legal compliance.

·Prepared by Convex Works from the European Software Data Study.

·Quick start

·import pandas as pd from datasets import load_dataset repo = "convexworks/cookie-banners" banners = load_dataset(repo, "banners", split="train").to_pandas() judgments = load_dataset(repo, "model_judgments", split="train").to_pandas() sites = load_dataset(repo, "sites", split="train").to_pandas() df = banners.merge(judgments, on="domain", validate="one_to_one").merge( sites[["domain", "vertical", "business_model"]], on="domain", validate="one_to_one" ) # One row per control, keeping the domain it belongs to. controls = pd.DataFrame( [{"domain": row.domain, "consent_tool": row.consent_tool, **c} for row in df.itertuples() for c in row.controls] ) buttons = controls[controls.kind == "button"].dropna(subset=["label"]) # Where the model found no direct reject button, what do the buttons say? no_reject = df.loc[df.reject_button == "not_detected", "domain"] print(buttons[buttons.domain.isin(no_reject)].label.str.casefold().value_counts().head(8))

·label accept 55 accept all 36 cookie settings 24 settings 16 accept all cookies 15 ok 12 learn more 11 customize 10

·Many sites use a consent tool's default wording: the 1,010 banners have 701 distinct wordings. Use text_id to count each wording once.

·Tables

·Every table joins on domain, the company's domain as listed in the source directory. Each table has one row per domain.

·Versions and segments

·Version 1.0 covers companies classified as B2B software (vertical = software, business_model = b2b or b2b2c). Later versions add other segments as new rows in the same tables, so code written against 1.0 keeps working. Filter on the segment columns in sites to keep a fixed population, and pin a version with load_dataset(repo, "banners", revision="v1.0.0").

·Fields

·banners

·Each control: Field | Description

·model_judgments

·Model mistral-medium-2604, prompt version privacy_banner_saved_html_core_v7; every answer had confidence "high". The model read the original extracted fragments, not the cleaned rows in banners. not_detected means not found among the captured labels, not that the site has no such button. No person reviewed these judgments.

·sites

·The segment columns come from gpt-5.4-nano, one pass over each company's homepage, about, pricing and demo text. They were not reviewed by a person. no_banner_found and unclear are the banner model's reading of the saved HTML. They do not show that the site had no banner.

·How it was made

• 1Pages. Homepages of 2,357 companies listed in the EU-Startups directory, classified as B2B software and with estimated search traffic, saved as rendered HTML by a headless browser on 21-22 June 2026. No page was fetched again for this release.
• 2Candidates. A script collected page fragments with cookie or consent wording, together with their controls.
• 3Banner or not. mistral-medium-2604 judged whether each homepage's fragments contained a sitewide cookie banner: 1,179 yes, 1,090 no, 88 unclear. Answers that failed a consistency check against the captured labels were asked again with high reasoning.
• 4First screen. For eight consent tools (Cookiebot, CookieYes, HubSpot, Complianz, Finsweet Cookie Consent, iubenda, OneTrust and CookieScript), the first screen is located with the tool's own markup. Later layers such as settings panels, vendor lists and detail tabs are left out. For Complianz, the hidden parts follow its default stylesheet. Other banners use the whole banner element.
• 5Cleaning. Elements the HTML marks as hidden (the hidden attribute, inline display: none or visibility: hidden, or the tool's hide class) and screen-reader-only text are left out. Whitespace is normalised; the wording is otherwise unchanged. Emails and phone numbers would be replaced with [email] and [phone]; none remained in the released rows.
• 6Selection. 1,010 of the 1,179 banners pass every check. The rules were set by inspecting examples. No person checked the released rows. Exclusion reason | Sites | Meaning

·Limitations

• ·Saved HTML is not a screen recording. It does not show what was visible, how prominent each button was, or what clicking did. Stylesheets loaded from separate files were not available.
• ·One capture per site, from one location, at one time.
• ·Version 1.0 covers B2B software companies listed in EU-Startups with estimated search traffic, not a sample of all websites.
• ·kind follows the HTML: accept or reject options built as links appear as link.
• ·Banner detection, the accept/reject judgments and the segment columns come from model passes and were not reviewed by a person. Thirteen homepages contain the banner markup of one of the eight known consent tools but were not marked as having a banner (7 no_banner_found, 6 unclear).
• ·The dataset is not evidence of legal compliance or non-compliance.

·Licence and rights

·Our selection, structure and labels are released under CC BY 4.0. The banner wording is quoted from the websites listed; rights in it remain with the site owners and consent-tool vendors, and CC BY 4.0 does not cover it.

·Corrections and removal

·Site owners can ask for a row to be corrected or removed by writing to hello@convex.works or opening a discussion in this dataset's Community tab. Changes are applied in the next version.

·Citation

·@misc{convexworks_cookie_banners_2026, title = {Cookie Banner Text and Controls}, author = {{Convex Works}}, year = {2026}, url = {https://huggingface.co/datasets/convexworks/cookie-banners}, note = {Version 1.0.0} }

·Version

1.0.01.0.0 (October 2026): first release, B2B software companies.

·Use this dataset

·Downloads last month 43

·Number of rows: 4,377 Total file size: 732 kB