PrincetonPLI/cc-centrality-scores

Common Crawl Host-Graph Centrality Scores

Precomputed host-level graph centrality scores for the WebGraphMix pipeline, from Hubs or Fringes? Pretraining Data Selection via Web Graph Centrality.

Scores are computed on the undirected host graph obtained by intersecting the Common Crawl host-level web graph with hosts present in WebOrganizer/Corpus-200B. They are used to importance-sample central ("hub") vs. peripheral ("fringe") documents for language model pretraining.

Resource	Link
Paper	arXiv:2606.11499
Project page	princeton-pli.github.io/WebGraphMix
Code	github.com/princeton-pli/WebGraphMix
1B checkpoints	PrincetonPLI/WebGraphMix-openlm-1B
Base corpus	WebOrganizer/Corpus-200B

Files

File	Metric	Size (approx.)	Hosts
host_graph_scores_betweenness_k1400000.json	Betweenness centrality	596 MB	13,906,789
host_graph_scores_eigenvector_maxiter1000.json	Eigenvector centrality	641 MB	13,906,789
host_graph_scores_katz_maxiter1000.json	Katz centrality	641 MB	13,906,789

Total download size: ~1.88 GB.

Format

Each file is a single JSON object mapping hostname → float score:

{
  "example.com": 1.2640900592941762e-07,
  "news.example.org": 0.003599027404561639,
  ...
}

• Keys: normalized hostnames (no www. prefix; port stripped).
• Values: non-negative centrality scores (higher = more central under that metric).

When aligning scores to document URLs, use the same hostname normalization as the WebGraphMix pipeline (urlparse(url).netloc, strip port, strip leading www.). Documents whose hosts are absent from the graph receive score 0.0.

How scores were computed

Graph construction (see pipeline/graph/ in the WebGraphMix repo):

1. Download Common Crawl host-graph shards.
2. Build corpus_host_graph_undirected.pkl — CC graph restricted to Corpus-200B hosts.
3. Run GPU centrality with RAPIDS cuGraph / Dask.

Metric	Algorithm	Parameters
Betweenness	Approximate betweenness centrality	k=1,400,000, normalized, random_state=42
Eigenvector	Power iteration	max_iter=1000, tol=1e-5
Katz	Katz centrality	max_iter=1000, tol=1e-5

Recomputing from scratch requires multi-GPU hardware; downloading this dataset is the recommended path.

Download

huggingface-cli download PrincetonPLI/cc-centrality-scores \
  --local-dir ./pipeline/graph/centrality/results \
  --repo-type dataset

Or from the WebGraphMix repo:

git clone https://github.com/princeton-pli/WebGraphMix.git
cd WebGraphMix
./experiments/artifacts/download.sh centrality

Usage in WebGraphMix

After download, scores plug directly into the annotation and sampling pipeline:

# 1. Generate per-document centrality annotations (betweenness example)
python pipeline/annotations/centrality_topk.py --centrality-metric betweenness

# 2. Importance-sample a 50/50 top/bottom mix
#    (configure OUTPUT_ROOT and other env vars — see run_sampling_job.sh)
./pipeline/sampling/run_sampling_job.sh

# 3. Tokenize, train, evaluate — see WebGraphMix README

Analysis utilities:

python pipeline/graph/centrality/analyze_scores.py   # score distributions, example hosts
python pipeline/graph/doc_counts_by_host.py          # docs/tokens by centrality percentile

Example: load scores in Python

import json

with open("host_graph_scores_betweenness_k1400000.json") as f:
    host_scores = json.load(f)

print(f"{len(host_scores):,} hosts")
print(host_scores["wikipedia.org"])  # if present in graph

Intended use

• Research: construct web-graph–aware pretraining mixtures (central vs. peripheral documents).
• Analysis: study how graph structure correlates with document/token counts in large web corpora.
• Extension: combine centrality with other signals (e.g. DCLM-fasttext quality scores) as in WebGraphMix+.

These files are not document-level labels; they are host-level scores mapped to documents via URL hostname at annotation time.

Citation

@article{badoni2026webgraphmix,
  title={Hubs or Fringes: Pretraining Data Selection via Web Graph Centrality},
  author={Badoni, Vedant and Chen, Danqi and Wang, Xinyi},
  year={2026}
}

License

Released under the MIT License, consistent with the WebGraphMix code release.