Pulpie has launched a family of Pareto-optimal models for extracting main content from HTML pages, achieving state-of-the-art (SOTA) quality at significantly reduced costs. Pulpie's smallest model processes 13.7 pages per second on an NVIDIA L4 GPU, compared to Dripper's 0.68 pages, making it a compelling tool for web content cleaning.
Pulpie has introduced a series of Pareto-optimal models designed for extracting main content from HTML pages. These models achieve quality comparable to existing state-of-the-art solutions such as Dripper but at a significantly lower operational cost.
The smallest model, pulpie-orange-small, has a ROUGE-5 F1 score of 0.862 on the WebMainBench, closely matching Dripper's performance of 0.864 while having only 210 million parameters compared to Dripper's 600 million.
In terms of processing speed, pulpie-orange-small can handle 13.7 pages per second on an NVIDIA L4 GPU, starkly outperforming Dripper, which processes 0.68 pages per second. This efficiency dramatically reduces the cost of cleaning large volumes of web pages.
The efficiency gains stem from Pulpie's architecture, where it labels HTML blocks as content or boilerplate in a single forward pass. This streamlines the extraction process and minimizes computational overhead. The estimated cost for cleaning 1 billion pages is $7,900 with Pulpie versus $159,000 with Dripper.
The initiative addresses the well-known issue of extraction being a bottleneck in language model training. By improving extraction quality, Pulpie enhances the quality of data consumed by language models during both pre-training and inference stages.
Research, such as the findings by Ma et al. in 2025, shows that cleaner extraction significantly boosts model performance across various benchmarks. The advantage of using models like Pulpie indicates the importance of clean data for training AI systems effectively.
Pulpie's models are open sourced and can be found on Hugging Face, providing developers and researchers with access to high-performance extraction tools. This opens up new opportunities for efficient data cleaning and enhances the overall quality of language model training.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
Pulpie has launched a family of Pareto-optimal models for extracting main content from HTML pages, achieving state-of-the-art (SOTA) quality at significantly reduced costs. Pulpie's smallest model processes 13.7 pages per second on an NVIDIA L4 GPU, compared to Dripper's 0.68 pages, making it a compelling tool for web content cleaning.