Web Graph Centrality Used to Optimize Language Model Pretraining Data Selection
Researchers propose WebGraphMix, a method that uses web graph structure to select pretraining data for language models by balancing central and peripheral web hosts. The approach requires no model training or labeled data, instead leveraging the Common Crawl web graph to compute efficiency scores at scale. The findings suggest that combining central hosts (which provide reusable abstractions) with peripheral hosts (which encode specialized knowledge) improves model performance across diverse tasks.
A new pretraining data selection framework called WebGraphMix uses structural centrality scores from the Common Crawl host-level web graph to optimize language model training. Rather than relying on auxiliary classifiers or mixture optimization methods that require labeled data and computational overhead, WebGraphMix computes centrality scores efficiently at web scale. The researchers trained models at 400M and 1B parameter scales and found that a 1:1 mixture of central and peripheral web regions achieved 41.4% average performance across 23 evaluation tasks, compared to 39.8% for uniform sampling. When combined with document-level quality classifiers, performance improved further to 43.8%. The work demonstrates that web graph topology captures information largely orthogonal to existing content-based data selection approaches, suggesting that structural properties of the web encode meaningful signals for model pretraining.
What different sources said
- arXiv cs.AICenter
Hubs or Fringes: Pretraining Data Selection via Web Graph Centrality
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.