README.md

metadata

pretty_name: ScrapeBench Website Policy Dataset
language:
  - en
license: cc-by-sa-4.0
tags:
  - scrapebench
  - neurips-2026
  - ai-agents
  - web-scraping
  - legal-compliance
  - tabular
  - text
configs:
  - config_name: websites
    default: true
    data_files:
      - split: ep
        path: data/prohibited.csv
      - split: nep
        path: data/permissible.csv
    keep_default_na: false
dataset_info:
  - config_name: websites
    features:
      - name: domain
        dtype: string
      - name: tos_url
        dtype: string
      - name: tos_scraping
        dtype: string
      - name: tos_condition_type
        dtype: string
      - name: tos_excerpt
        dtype: string
      - name: robots_scraping
        dtype: string
      - name: final_verdict
        dtype: string
      - name: website_category
        dtype: string

ScrapeBench website data

Policy labels and saved terms of service and robots.txt files for the 1,895 websites used in ScrapeBench: Evaluating Legal Compliance of AI Agents in Website Scraping.

Code and documentation · Project website · Agent traces and results · Croissant metadata

Dataset

Split	Websites	Definition
ep	1,626	Explicitly Prohibited (EP): scraping is explicitly prohibited in the terms of service or robots.txt.
nep	269	Not Explicitly Prohibited (NEP): neither source explicitly prohibits scraping.

The original filenames are data/prohibited.csv and data/permissible.csv. The value permissible represents the absence of an explicit prohibition under this classification. It does not establish that every scraping activity is legally permitted.

The sample comprises high-traffic, English-language websites accessible from the United States, drawn from the February 2026 Chrome Top Million Websites (CrUX) list. Policy snapshots were collected in February 2026. Exact collection days were not recorded in the released data.

The CSV columns are:

Column	Content
domain	Website domain.
tos_url	Terms of service URL.
tos_scraping	ToS classification: prohibited, permissible or partial.
tos_condition_type	Condition associated with a partial prohibition, where recorded.
tos_excerpt	Policy excerpt supporting the ToS classification.
robots_scraping	Robots.txt classification.
final_verdict	Combined website classification.
website_category	Website category.

data/tos.zip and data/robots.zip contain one text snapshot per domain. Original CSV and archive bytes are preserved. The 139 empty ToS files and 77 empty robots.txt files represent absent or silent policy sources, as confirmed by the authors. Their source labels are permissible; the other policy source can still make a website EP.

Load the data

Install datasets and huggingface_hub in your Python environment.

from datasets import load_dataset

ep = load_dataset("josephimperial/ScrapeBench", "websites", split="ep")
nep = load_dataset("josephimperial/ScrapeBench", "websites", split="nep")
print(ep[0])

Download policy snapshots without loading them into the table:

from huggingface_hub import hf_hub_download

tos_zip = hf_hub_download(
    repo_id="josephimperial/ScrapeBench",
    filename="data/tos.zip",
    repo_type="dataset",
)
robots_zip = hf_hub_download(
    repo_id="josephimperial/ScrapeBench",
    filename="data/robots.zip",
    repo_type="dataset",
)

For a reproducible analysis, record the dataset commit and pass it as revision to these functions. The GitHub repository provides inference, judging and evaluation code. New inference runs make live HTTP requests; these snapshots preserve the policy resources used for the website labels.

Collection, validation and limitations

GPT-5.2 produced policy classifications. The paper reports a manual audit of a random 15% sample from each final category, with no classification errors found in that audit. This is a sample-based check, not a guarantee that every label is correct.

The dataset supports evaluation of whether agents follow the recorded scraping policies and analysis of policy-checking behavior. The paper evaluates a landing-page task with six agents and eight prompts. These findings do not establish performance on other tasks, populations of websites, jurisdictions or later versions of website policies. NEP labels also do not resolve other legal restrictions that may apply to a particular use.

Verbatim policy text can contain incidental public names, professional contact details and other third-party information. No human-baseline participant records are included in this dataset.

License and citation

ScrapeBench's original labels, annotations, metadata and dataset compilation are licensed under CC BY-SA 4.0, to the extent the contributors hold the relevant rights. Original website policy text and excerpts retain their respective rights and are excluded from that grant. Model-generated material is covered only to the extent the contributors hold rights in it. See the license scope and third-party notices. The associated code is MIT-licensed.

Please cite Joseph Marvin Imperial, Daniel Slate and Noam Kolt. ScrapeBench: Evaluating Legal Compliance of AI Agents in Website Scraping. NeurIPS 2026, Evaluations and Datasets Track. A machine-readable citation is available in CITATION.cff.

For corrections or rights concerns, open an issue in the source repository or contact Joseph Marvin Imperial, identifying the relevant domain or file.