HalfLife and Web-Scale Data Poisoning

A concise, abstract-grounded briefing on a paper that introduces HalfLife to estimate whether adversarial content from public discussion interfaces enters web-crawl based LM training data.

TL;DR — The abstract argues that pretraining-data poisoning is feasible beyond Wikipedia-like sources. It identifies public discussion interfaces as a web-scale content injection mechanism and introduces HalfLife to estimate whether adversarial content is included in web-crawl based LM training data after crawling and curation.

Problem and scope

The abstract frames the paper around a specific security gap: poisoned pretraining data can produce harmful language-model behaviors that are difficult to detect and mitigate. The key scope decision is to move beyond prior work centered on established data sources such as Wikipedia and toward the scale and heterogeneity of web-crawl based pretraining corpora.

Full analysis, extracted claims, numerical results, entity graph, FAQ, related work, applications, and BibTeX are available via x402 micropayment.