Hey HN, I'm Shreyash, founder of Feyn. We built Pulpie, a family of Pareto optimal models for cleaning the web. Pulpie strips boilerplate (ads, footers, sidebars) from raw HTML and returns just the main content as HTML or Markdown.
We match SOTA extraction quality while being 20x cheaper. Cleaning 1 billion webpages costs $7,900 with Pulpie versus $159,000 with Dripper, the current leading extractor.
The gains come from architecture. Today's leading extractors are decoders that generate output one token at a time. Each step reads the full model from memory to produce a single token. Conversely, Pulpie models are encoders. They run one forward pass over the full input HTML and label each block as boilerplate or content. As a result, Pulpie is compute-bound while decoders are memory-bound. Cheaper GPUs have relatively more compute than memory bandwidth. This makes Pulpie easy to run optimally.
Our motivation for Pulpie came from building a deep research harness. Every search API returns noisy content that contains ads, nav elements, and sidebars. In one instance, an ad for "Gemini on Pixel" slipped into our search results, got passed into LLM context, and ended up in the final answer served to the user. Pretty embarrassing moment for us but it helped us realize how bad data kills model intelligence. We built Pulpie to get clean data for cheap.
It's good looking, and I liked it. The trial page accessed from the hugging face website is a very inefficient experience when I use Mozilla and the dark theme, FYI.
Why does the 'Quality vs Cost of Web Content Extraction' chart not have zero cost at the origin? Up to the right does not have to mean better; we can read.
If I had to reckon, it's because the web comes in very many shapes, and outsourcing that work to a generalist LLM/SLM like GPT Nano is expensive, and doing it deterministically will never catch all the edge cases as well as a purpose-built encoder when run at webscale.
Looks like they are including Trafilatura in the comparison tables, which I've used before with pretty decent results, but it still has trouble with some pages. Looks like the pulpie f1 scores are quite a bit better, especially for the hard cases.
Would be curious how it runs on more modest hardware though, I'm using it for a small bookmark archiving tool and being able to run it on my small mini-pc homelab would be nice.
We see far better performance with models. Heuristics break on richer content like codeblocks, formulae, quotes, etc. In our testing, our model was 25 F1 points better than Trafilatura.
I think instead of "performance" you must mean "accuracy". Traditional deterministic conversion will always be faster and cheaper than running through a model, even if it is less accurate.
So this is tailored towards kind of a "reader view" for models right? Can it handle images, tables, shadow DOMs too? Like there are 3 use cases I have now - one is a simple text view for models to understand it, one is a "web clip" mode which would ideally preserve images and media, and one is to extract tabular data from web pages. Which ones is this good at?
I did some research on this about 10 years ago. I spent 2 days hand labelling data from scraped news sites. Then built a good old fashioned Random Forest model to classify html nodes based on some feature engineering. turns out the P tag and the number-of-words threshold get you 90% of the way there, on news sites anyway. Great thing about RF models is they tell you which features are the most important. fun little project (apart from the 2 days of data labelling).
35 comments
[ 3.0 ms ] story [ 40.9 ms ] threadWe match SOTA extraction quality while being 20x cheaper. Cleaning 1 billion webpages costs $7,900 with Pulpie versus $159,000 with Dripper, the current leading extractor.
The gains come from architecture. Today's leading extractors are decoders that generate output one token at a time. Each step reads the full model from memory to produce a single token. Conversely, Pulpie models are encoders. They run one forward pass over the full input HTML and label each block as boilerplate or content. As a result, Pulpie is compute-bound while decoders are memory-bound. Cheaper GPUs have relatively more compute than memory bandwidth. This makes Pulpie easy to run optimally.
Here's Pulpie and Dripper cleaning the same pages side by side: https://www.youtube.com/watch?v=ibd-tIiQECo. You can try a side-by-side comparison yourself: https://huggingface.co/spaces/feyninc/pulpie
Our motivation for Pulpie came from building a deep research harness. Every search API returns noisy content that contains ads, nav elements, and sidebars. In one instance, an ad for "Gemini on Pixel" slipped into our search results, got passed into LLM context, and ended up in the final answer served to the user. Pretty embarrassing moment for us but it helped us realize how bad data kills model intelligence. We built Pulpie to get clean data for cheap.
All models are open source on Hugging Face. You can read about our training process and how to use Pulpie here: https://usefeyn.com/blog/pulpie-pareto-optimal-models-for-cl...
Happy to answer any questions!
Would be curious how it runs on more modest hardware though, I'm using it for a small bookmark archiving tool and being able to run it on my small mini-pc homelab would be nice.
HF Space: https://huggingface.co/spaces/feyninc/pulpie