- Pulpie is a family of Pareto-optimal models for extracting main content from HTML pages, achieving near-state-of-the-art quality at significantly lower cost.
- The smallest model, Pulpie Orange Small (210M parameters), matches Dripper's extraction quality (0.862 vs 0.864 ROUGE-5 F1) but is three times smaller and much faster (13.7 vs 0.68 pages/sec on L4 GPU).
- Cost savings are substantial: cleaning 1 billion pages costs $7,900 with Pulpie vs $159,000 with Dripper on an L4 GPU.
- Pulpie uses an encoder architecture that labels all HTML blocks in one forward pass, making it compute-bound and faster than decoder-based extractors like Dripper, which are memory-bandwidth-bound.
- Training involved creating a dataset from Common Crawl, labeling blocks with DeepSeek V3.2, and distilling a 2.1B teacher model into smaller models (210M and 610M) via knowledge distillation.
- Pulpie performs well across difficulty levels, with heuristics declining more on hard pages, and handles long pages better by chunking blocks into 8,192-token segments.
- The models are open-source on Hugging Face, with easy installation and usage for both single-page and batch processing, recommending Pulpie Orange Small for production.
- Clean extraction improves both pre-training (e.g., +1.08% accuracy in AICC study) and inference by reducing noise, highlighting the importance of high-quality web data.