Pulpie Offers Pareto Optimal Models for Cleaning the Web

Pulpie is a family of encoder models built to mark HTML blocks as content or template. The smallest version reaches near parity with much larger tools while using far less compute.

Pulpie-orange-small uses 210 million parameters. It scores 0.862 ROUGE-5 F1 on WebMainBench. That result sits one point behind the 600 million parameter Dripper model.

The speed gap is larger. Pulpie processes 13.7 pages per second on an NVIDIA L4 GPU. Dripper manages 0.68 pages per second on the same hardware.

Cost numbers follow the same pattern. Cleaning one billion HTML pages costs roughly 7900 dollars with Pulpie. The same task costs about 159000 dollars with Dripper.

The models use a single forward pass to label every block. No extra stages or heavy post-processing are required.

Smaller Size Matches Larger Accuracy

Pulpie-orange-small closes most of the accuracy gap to Dripper despite having one third the parameters. The result shows that Pareto optimal choices can keep quality high without scaling size.

The encoder architecture keeps inference simple. Each HTML chunk receives one label in a single pass. This design removes sequential steps that slow down other extractors.

Speed Advantage Changes Economics

The 20 times throughput gain comes from the compact encoder design. Fewer layers and smaller hidden states reduce both memory traffic and compute time.

Faster inference directly lowers cloud bills. Teams that clean large crawls see the biggest savings when page counts reach hundreds of millions or more.

Open Release Lowers Adoption Barriers

Both models sit on Hugging Face. Developers can download weights and run local tests without signing license agreements. Open weights let researchers measure performance on their own datasets rather than rely on vendor benchmarks.

Limits Still Require Human Checks

Pulpie focuses on content versus template labels. It does not resolve nested scripts or handle heavy JavaScript rendering. Sites that load main text after initial HTML may need extra steps before Pulpie sees the full page.

The published scores come from WebMainBench. Results on other corpora can shift. Teams should run small pilots on their target domains before committing large jobs.

Next Signals to Track

Watch release notes for updates to the model family. New sizes or fine-tunes may appear in the coming months.

Track independent benchmark submissions on WebMainBench and similar suites. Consistent scores above 0.86 with parameter counts under 300 million would strengthen the Pareto claim.

Observe adoption in open source crawlers. Usage numbers on GitHub or public pipeline forks will show whether the cost and speed gains move from benchmark to production.

Pulpie shows that targeted architectural choices can shift the cost curve for web cleaning. Developers who run repeated large crawls now have a documented lighter option that keeps most accuracy. Continued measurement on fresh data will confirm how far the gains extend beyond the initial test set.