Dripper: Token-Efficient Main HTML Extraction with a Lightweight LM

High-quality main content extraction from web pages is a critical prerequisite for constructing large-scale training corpora. While traditional heuristic extractors are efficient, they lack the semantic reasoning required to handle the structural heterogeneity of the modern web. Conversely, well-pretrained generative Large Language Models (LLMs) offer superior document comprehension but are prohibited by excessive computational costs, limited context windows, and hallucination risks when applied at web scale. We present $\mathbf{\text{Dripper}}$, a lightweight framework that resolves these bottlenecks through four contributions: (1) We reformulate extraction as a $\mathbf{\text{constrained}}\mathbf{\text{sequence}}\mathbf{\text{labeling}}$ task using SLMs (Small Language Models). This paradigm eliminates generative hallucinations and achieves exceptional efficiency, reaching a throughput of 3.08 pages per second on a single A100 GPU. (2) We construct $\mathbf{\text{WebMainBench}}$, a rigorous benchmark of 7,809 human-annotated pages covering 5,434 unique domains and multiple languages. Evaluations show our Dripper-0.6B model $\mathbf{\text{outperforms}}$ heuristics like Trafilatura and rivals massive models like DeepSeek-V3.2(685B), GPT-5 and Gemini-2.5-Pro, offering an optimal efficiency-accuracy trade-off. (3) We demonstrate infrastructural value by $\mathbf{\text{pre-training}}\mathbf{\text{a}}\mathbf{\text{1B}}\mathbf{\text{model}}$ on a Dripper-curated corpus (63B tokens). This model significantly outperforms baselines in downstream tasks, proving the critical role of extraction quality and the effectiveness of our framework. (4) We $\mathbf{\text{open-source}}$ the Dripper-0.6B weights and codebase to facilitate the construction of high-quality datasets.