Establish whether crawl-yield gains improve trained-model quality

Determine whether the higher high-quality-page yield achieved by uniform, centrality-based, and sample-then-deepen crawling policies translates into improved downstream language-model performance.

Background

The paper measures the number and density of classifier-selected high-quality pages collected by alternative crawl policies, rather than training LLMs on the resulting datasets. Because classifier-based proxy metrics may diverge from downstream accuracy, the relationship between crawl yield and actual model quality remains unresolved. The proposed experiment would train small models on data selected by the competing policies and compare their performance.

References

Downstream model ablation. We measure crawl yield and coverage, not trained-model performance. Given that proxy metrics and downstream accuracy can diverge \citep{li2024datacomp,yu2025craw4llm}, one would need to train small models ($\sim$1B parameters) on data selected by uniform, centrality, and sample-then-deepen policies test whether the yield gains translate into downstream quality.

— Host-Aware Web Crawling for LLM Pretraining Data  (2610.13321 - Ostendorff, 9 Oct 2026) in Section Future Work, subsection “Downstream model ablation”