Gemini 4 Carbon: Our next frontier model
deepmind.google·5d ago
AI Engineering
Hacker News·Jul 6·5 min read
Show HN: Pulpie – Models for Cleaning the Web
TL;DR
Pulpie introduces a family of Pareto-optimal models for extracting main content from HTML pages, achieving state-of-the-art quality at a fraction of the cost. The smallest model, pulpie-orange-small, matches the performance of the leading extractor Dripper while being significantly smaller and faster.
✦ Why It Matters
Engineers can adopt Pulpie for cost-effective web content extraction, significantly reducing operational expenses.
Key Takeaways
- Implement Pulpie for efficient web content extraction.
- Utilize pulpie-orange-small for cost-effective processing.
- Leverage the architecture for faster HTML block labeling.
- Reduce operational costs significantly when cleaning large datasets.
Full Summary
Web content extraction is often hindered by high costs and inefficiencies. Pulpie is a new family of models designed to extract main content from HTML pages using a Pareto-optimal approach, which balances quality and cost.
The smallest model, pulpie-orange-small, has 210 million parameters and achieves a ROUGE-5 F1 score of 0.862, closely matching the leading extractor Dripper, which has 600 million parameters and a score of 0.864. Pulpie's architecture allows it to label HTML blocks as content or boilerplate in a single forward pass, resulting in faster processing speeds.
On an NVIDIA L4 GPU, pulpie-orange-small processes 13.7 pages per second, compared to Dripper's 0.68 pages per second. This efficiency translates to a cost of $7,900 for cleaning 1 billion pages with Pulpie, versus $159,000 with Dripper.
These advancements are expected to enhance pre-training and context management in various applications.