Evaluating focused web crawlers on the live internet is notoriously messy: remote websites change constantly, servers throttle requests unpredictably, and testing a crawler against its own internal classifier is an evaluation trap.

ProdCrawl fixes this by providing an offline, deterministic playground. It contains complete static mirrors of 621 company and organizational websites (over 1.1 million pages) paired with precomputed ground-truth labels for every single reachable page. Because every crawl replays identical web pages against identical target labels, different link prioritization strategies can finally be compared under strictly controlled conditions.

Code & Reproduction Framework: github.com/IRLab-UDC/prodcrawl
Fine-tuned Bi-Encoder: irlab-udc/sbert-anchor-product-page-v5

Dataset Overview

The collection is organized into relevance judgments (for scoring) and compressed web archives (for serving the local web environment):

judgments/{train,validation,test}.tsv        Relevance judgment for every reachable page
warcs/{train,validation,test}/<site>.warc.gz One compressed WARC archive per domain
warcs/manifest.tsv                           Manifest with paths, record counts, and SHA-256 hashes
warcs/manifests/<site>.json                  Per-site mirroring metadata

Splits

All splits are partitioned strictly by website domain. Web pages within the same domain share navigation menus, boilerplate, and anchor text phrasing; splitting at the page level would cause severe data leakage.

Split	Sites	Total Pages	Product Pages	WARC Size
Train	318	544,059	208,124 (38.3%)	15.0 GB
Validation	167	301,312	72,847 (24.2%)	7.2 GB
Test	136	257,563	116,480 (45.2%)	7.0 GB
Total	621	1,102,934	397,451 (36.0%)	29.2 GB

Every domain contains at least one target product page, and English is the primary language across ~95% of the sites.

The Crawling Task

The crawler begins at a domain's homepage with a constrained download budget and must prioritize which links to follow to discover high-value transactional product pages as early as possible.

What Counts as a Product Page?

Relevant (True): A page whose clear goal is bottom-of-the-funnel conversion for a specific item, service tier, or software solution. Hallmarks include pricing tables, feature specifications, SKU descriptions, demo request forms, or "Add to Cart" actions.
Not Relevant (False): Top- or mid-funnel content such as corporate overviews, blogs, news announcements, knowledge base articles, general documentation, or broad navigation hubs lacking a direct conversion mechanism.

Relevance Judgments & Ground Truth

To avoid manual labeling bottlenecks across 1.1M documents while eliminating circular reliance on crawler-internal models, ground truth was established using an LLM jury:

Judges: Llama 3.1 8B Instruct, Llama 4 Scout 17B-16E Instruct, and Gemma 3 27B IT.
Protocol: Text was extracted with ReadabiliPy (first 10,000 characters) and judged independently via constrained greedy decoding (Yes/No) under an identical task prompt.
Consensus: Final labels represent majority consensus. Across all splits, 84% to 89% of decisions were unanimous, with Fleiss' $\kappa$ between 0.76 and 0.85 (substantial agreement).

Loading Judgments

from datasets import load_dataset

judgments = load_dataset("irlab-udc/prodcrawl", "judgments")
print(judgments["test"][0])
# Example output: {'url': 'http://www.example.com/products/tool.html', 'label': True}

Web Mirrors & Replay Architecture

Each website is stored as an individual gzipped WARC/1.1 file (warcs/<split>/<site>.warc.gz) containing the deterministic HTML responses captured via HTTrack.

Reading WARCs Programmatically

from warcio.archiveiterator import ArchiveIterator

with open("warcs/test/guesttek.com.warc.gz", "rb") as f:
    for record in ArchiveIterator(f):
        if record.rec_type == "resource":
            target_url = record.rec_headers.get_header("WARC-Target-URI")
            raw_html = record.content_stream().read()

Downloading Specific Splits via Hugging Face Hub

from huggingface_hub import snapshot_download

# Download only the test split (WARCs + ground-truth TSV)
snapshot_download(
    repo_id="irlab-udc/prodcrawl",
    repo_type="dataset",
    allow_patterns=["judgments/test.tsv", "warcs/test/*"]
)

URL Structure Notes

The mirror URLs reflect offline archive paths rather than live endpoints:

Query strings are encoded directly into path segments (e.g., [http://example.com/__querystring__p=12/index_no_slash.html](http://example.com/__querystring__p=12/index_no_slash.html)).
Missing file extensions or root directories typically map to .../index.html or .../index_no_slash.html.
All URLs are normalized to http:// using the site's original hostname.

How to Run & Evaluate a Crawler

Host the Mirrors: Unpack the desired WARC split and serve it locally using Apache or Nginx as detailed in the reproduction repository.
Execute Traversal: Point your crawler at http://<local-server>/<site-domain>/ with your link prioritization policy active.
Score the Trace: After running for a budget of $b$ requests, evaluate the ordered sequence of fetched URLs against judgments/<split>.tsv. Any visited page not found in the ground truth (e.g., auto-generated directory listings) is treated as non-relevant (False).

The standard metric is Mean Average Precision normalized by domain targets and the Area Under the $M A P$ Curve for early traversal budgets (e.g., $b \in [1, 30]$ ):

$MAP@b = \frac{1}{\vert{}D\vert{}} \sum_{d \in D} \frac{1}{R_d} \sum_{i \le b, \, rel(p_i^d)=1} \text{Precision}@i$

where $R_{d}$ is the total count of relevant product pages in domain $d$ .

License & Fair Use

The judgments, metadata, and collection structure are provided under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

The underlying web documents were gathered and archived strictly for non-commercial academic research, empirical reproducibility, and information retrieval evaluation under fair use principles. Copyright for individual website contents remains with the original domain publishers. To request removal of a mirrored domain, open an issue in the Community tab.

Citation

@misc{reyescobar2026prodcrawl,
  title        = {ProdCrawl: A Replayable Test Collection for Focused Crawling},
  author       = {Rey-Escobar, Manuel and Otero, David and Parapar, Javier},
  year         = {2026},
  howpublished = {\url{https://github.com/IRLab-UDC/prodcrawl}},
  note         = {Dataset available at Hugging Face: \url{https://huggingface.co/datasets/irlab-udc/prodcrawl}}
}

Manuel Rey-Escobar, David Otero, and Javier Parapar. 2026.
ProdCrawl: A Replayable Test Collection for Focused Crawling.
Available at https://github.com/IRLab-UDC/prodcrawl

Downloads last month: 21

Total file size:

29.3 GB