Show HN: Pulpie – A Model for Web Content Cleaning

A Pareto-optimal model family called Pulpie, which removes unnecessary elements like ads and navigation from web page HTML structures and extracts only the main content, has been released.

Compared to the existing SOTA model Dripper, Pulpie has the following technical advantages.

• High efficiency: Despite being 210M parameters in scale, 3 times smaller than Dripper, it recorded comparable performance on WebMainBench.
• Overwhelming speed: Using an Encoder architecture, it labels all blocks in a single Forward pass, processing 13.7 pages per second on an NVIDIA L4 GPU (about 20x faster than Dripper).
• Dramatic cost reduction: The cost of cleaning 1 billion pages is around $7,900, about 20 times cheaper compared to Dripper ($159,000).

Cleaned data plays a key role in improving model accuracy during the LLM's Pre-training stage, and in maximizing model efficiency by reducing context noise during the Inference stage. The Pulpie model has been open-sourced via Hugging Face.