Beyond a Single Extractor: Re-thinking HTML-to-Text Extraction for LLM Pretraining
This addresses a data preprocessing bottleneck for LLM developers, offering a simple intervention to improve dataset coverage and downstream performance, though it is incremental as it builds on existing extraction methods.
The paper tackled the problem of suboptimal text extraction from HTML for LLM pretraining by showing that using a single extractor reduces coverage, and found that taking a union over different extractors increased token yield by up to 71% while maintaining performance, with extractor choice impacting downstream tasks by up to 10 percentage points on WikiTQ and 3 percentage points on HumanEval.
One of the first pre-processing steps for constructing web-scale LLM pretraining datasets involves extracting text from HTML. Despite the immense diversity of web content, existing open-source datasets predominantly apply a single fixed extractor to all webpages. In this work, we investigate whether this practice leads to suboptimal coverage and utilization of Internet data. We first show that while different extractors may lead to similar model performance on standard language understanding tasks, the pages surviving a fixed filtering pipeline can differ substantially. This suggests a simple intervention: by taking a Union over different extractors, we can increase the token yield of DCLM-Baseline by up to 71% while maintaining benchmark performance. We further show that for structured content such as tables and code blocks, extractor choice can significantly impact downstream task performance, with differences of up to 10 percentage points (p.p.) on WikiTQ and 3 p.p. on HumanEval.