You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

GlotCrawl-60 holds text of public web pages in 60 low-resource languages. Access is for non-commercial research only. To get access, fill in the form and accept the research-only data licence (LICENSE-data). Access is granted when you submit the form.

Log in or Sign Up to review the conditions and access this dataset content.

GlotCrawl-60

GlotCrawl-60 is a closed copy of the web around the seed sites of 60 low-resource languages, made for the study of focused crawlers. It holds every page that a crawler can fetch from the seed sites and one link away, the links between them, the page texts and link texts, and the outputs of the study runs. A crawler can run on this copy again with no web access, and it gets the same pages in the same order.

Access. The whole dataset is behind a gate. Fill in the form and accept the research-only data licence (LICENSE-data). Access is granted when you submit the form. Use the data only for non-commercial research. Do not redistribute the page texts or link texts.

Splits

  • dev: 20 languages. Use them to develop and tune a crawler.
  • test: 40 languages. Use them only for the final evaluation. The study reports results on these.

Each Parquet file holds one language (<config>/<split>/<label>.parquet, column label, for example lld_Latn). You can download one language alone. Optional configs (scores_jev, scores_pl, html) exist only where the study has them.

Config Split Languages Rows Size
pages dev 20 1,202,423 67.6 MB
pages test 40 2,640,481 151.4 MB
page_texts dev 20 608,622 762.9 MB
page_texts test 40 1,347,805 1.4 GB
links dev 20 75,004,177 687.4 MB
links test 40 182,323,401 1.5 GB
link_texts dev 20 75,004,177 1.3 GB
link_texts test 40 182,323,401 3.0 GB
robots dev 20 104,151 121.6 MB
robots test 40 227,742 313.8 MB
onehop dev 20 7,302,610 170.2 MB
onehop test 40 15,074,457 340.1 MB
seeds dev 20 538 44.7 KB
seeds test 40 1,020 86.0 KB
label_info dev 20 20 282.1 KB
label_info test 40 40 580.9 KB
run_summaries dev 20 20 362.1 KB
run_summaries test 40 290 1.7 MB
fetches dev 20 57,314 3.7 MB
fetches test 40 990,192 57.7 MB
scores_jev test 40 1,162,672 26.8 MB
scores_pl test 10 68,206,427 153.0 MB

Configs

  • pages: One row per stored URL: status, redirect target, content type, sizes, time, GlotLID label and confidence of the page text, text SHA-1, filter flags, number of links (rid = row id).
  • page_texts: Main text of each stored page with status < 400, as trafilatura.extract gives it (the same call that made the page labels). sha1_match is true if the SHA-1 equals pages.text_sha1.
  • links: One row per <a href> of each page the crawler analyzes, in document order (rid, pos), with GlotLID numbers about the anchor and context text.
  • link_texts: Anchor text, 100 characters of text left and right of the link, and the link attributes, for the same rows as links (join on label, rid, pos).
  • robots: The robots.txt file of each site as the crawl stored it.
  • onehop: URLs one link away from the seed sites (candidates of the closed copy).
  • seeds: Seed URLs of each language, in file order.
  • label_info: One row per language: site information and build information (JSON).
  • run_summaries: One row per study run (crawler and language): budget, fetches, final number of verified pages in L, curve, digest, full summary (JSON).
  • fetches: One row per fetch of each study run, in fetch order, with the score that ordered the frontier.
  • scores_jev: Cached link probabilities of the Jev scorer (TypeSafe, model jev-1.14.0), keyed by the cache key of src/jev_scorer.py. A replay with this cache needs no API call.
  • scores_pl: Per-link score of the per-language fastText scorer (crawler trained-per-language, PL), for the languages that have one.

Other folders: runs/<split>/<label>/<run>/replay_requests.jsonl.gz (the request log of each study run, for the replay check), config/ (study configuration), results/ (study results), models/ (the link scorers of the study, see below), code/ (the converter convert_hf_download.py), checksums/ (SHA-256 of every file). The files in config/, results/ and code/ have the same names and content as in the code repository (code/convert_hf_download.py is scripts/convert_hf_download.py there). results/runs/ holds one run summary per study run set (for example main_runs.json, rules.json), results/dev_split_runs/ the runs on the dev split.

Models

  • models/zs_final.pkl: the zero-shot link scorer of the crawlers zero-shot-fixed (ZS) and zero-shot-learning (ZSL) (scikit-learn HistGradientBoostingClassifier). It is a Python pickle: load it only if you trust this source. models/zs_final.json gives its training information.
  • models/zs_final_train.npz: the training matrix of the zero-shot scorer: X (float32, one row per link, one column per link feature) and y (int8, 1 = the link goes to a page in the language). It holds only numbers, no text. Load it with numpy.load(path, allow_pickle=False).
  • models/pl/pl_<label>_F2.bin: the per-language fastText link scorer of the crawler trained-per-language (PL), for 10 test languages (ast_Latn, bew_Latn, ceb_Latn, cos_Latn, dsb_Latn, hsb_Latn, lld_Latn, ltg_Latn, pfl_Latn, rmn_Latn). Each model has a .json file with its training information. A fastText model holds the words and character n-grams of its training texts (link texts).
  • models/manifest.json: file, size and SHA-256 of each model file.

Load

from datasets import load_dataset
pages = load_dataset("anonymousaccount/GlotCrawl-60", "pages", split="test")                # needs a token with access
texts = load_dataset("anonymousaccount/GlotCrawl-60", "page_texts", data_files="page_texts/test/lld_Latn.parquet")

Replay a crawl with the code

  1. Get the code (https://anonymous.4open.science/r/wsdm2027-s1models-lrl-crawling-D716/) and install requirements.txt.
  2. Get the GlotLID model model_v3.bin from https://huggingface.co/cis-lmu/glotlid and set LRL_GLOTLID_MODEL to the file. The model is not part of this dataset.
  3. Download the configs pages, links, robots, onehop, seeds and label_info of one language.
  4. Convert them to the layout of the replay code: python code/convert_hf_download.py --hf-root <download dir> --split test --lang lld_Latn --out release_dir
  5. Run the replay: python scripts/replay.py --release-dir release_dir --lang lld_Latn --arm B ... (see docs/reproduce.md of the code). The run digest must equal run_summaries.digest. One command of the code does steps 4 and 5: python reproduce.py replay --lang lld_Latn --hf-root <download dir>.

These crawlers replay with the same digest: bfs (B), oracle (ORACLE), page-language-rule (R1), address-code-rule (R2), zero-shot-fixed (ZS) and zero-shot-learning (ZSL). ZS and ZSL use models/zs_final.pkl. The crawlers jev (JEV) and trained-per-language (PL) do not replay directly: scores_jev and scores_pl hold the link scores of the study runs, and models/pl/pl_<label>_F2.bin is the per-language model of PL. To replay PL, make a score table for --score-table from scores_pl (see DATA.md of the code). A JEV replay is not checked. The crawler decider (OPENDM) does not replay from the dataset.

Datasheet

  1. Collection. The sites come from a sample of hosts of GlotCC (derived from Common Crawl). They do not come from GlotWeb or from a search engine. A host of a test language is not used for a dev language.
  2. Crawl. User agent LRLCrawlResearch/0.1 (<e-mail of the authors>) (the e-mail is withheld for the anonymous review; in copied files <CONTACT> or <your e-mail> marks a removed e-mail address). The crawler respected robots.txt for this agent (cached for 24 hours), waited at least 1 s between two requests to one host and did not crawl blocklist domains. The crawl ran in November 2026.
  3. What is in the dataset. URLs, status codes, page labels, page texts (extracted main text), link texts (anchor and 100 characters of context), link numbers, robots.txt files, the study run outputs and the link scorers of the study (models/). Raw HTML is not included. No WARC files are included.
  4. robots.txt. The pages.pf_ccbot flag marks pages that a site blocks for CCBot. Use the flag to remove these pages if your use needs it.
  5. Copyright. The page texts belong to the site owners. The dataset gives them for research on crawling and language identification only, under the research-only licence. Do not redistribute them.
  6. Personal data. Web pages and URLs can contain names, e-mail addresses or ids. We did not remove them. Do not try to identify persons. See the takedown process below.
  7. Scope. The dataset is a closed copy of the web at crawl time. The pages may have changed or been removed since then. A crawler that needs a page outside the copy cannot get it.
  8. Bias. The languages come from GlotCC. Sites with many pages in a language are more likely to be in the sample. Most of the languages use Latin script.
  9. Jev scores. scores_jev holds probabilities from a paid API (model jev-1.14.0). A new run with an empty cache can give other values.
  10. Labels. The page labels come from GlotLID v3 on the extracted text. They are not human labels.
  11. Link extraction limit. The link extraction of one page stops after 300 s. Then pages.link_err is AnchorTimeout, the page row stays, and the page has no rows in links and link_texts. The build information in label_info lists these pages (anchor_timeout_pages).

Takedown

If you own a page in this dataset or a page names you, open a discussion in the Community tab of this dataset page. Give only the URL, no personal details. We remove the page from the dataset within 30 days and publish a new version. Users accept (checkbox) that they delete removed data from their copies.

Licence

  • Data: research-only licence, LICENSE-data (version 1.0).
  • Code in code/ and the configuration in config/: MIT (LICENSE).
  • Third-party parts keep their own licences (see below).

Citation

@misc{glotcrawl60,
  title     = {Decision Models in Web Crawling: A Case Study on Low-Resource Languages},
  author    = {Anonymous},
  note      = {Under review},
  year      = {2027}
}

Third-party notices

  • Common Crawl and GlotCC (Kargaran et al., NeurIPS 2024 Datasets and Benchmarks, https://huggingface.co/datasets/cis-lmu/GlotCC-V1): source of the sampled hosts and seed URLs. No GlotCC text is included. See the terms of use of Common Crawl and the licence of GlotCC.
  • GlotLID v3 (Kargaran et al., Findings of EMNLP 2023, https://huggingface.co/cis-lmu/glotlid): page labels and link numbers. The model is not included.
  • fastText (Joulin et al., EACL 2017, MIT licence): library of the per-language scorer.
  • trafilatura (Barbaresi, ACL 2021 System Demonstrations, Apache-2.0): main-text extraction.
  • GlotWeb blocklist (config/filterlist.txt, source in config/filterlist.SOURCE): see the licence of that repository.
  • The web pages belong to their owners.
Downloads last month
3,721