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.
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.
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.
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.
ongrats on the release! The architectural insight to switch from a bandwidth-bound decoder (token-by-token generation) to a compute-bound encoder (single forward pass over 8k chunks) is brilliant—the 20x speedup and cost drop from $159k to $7.9k per billion pages is massive for web-scale pipelines.
As someone building AI developer utilities and document tools, I have two quick technical questions:
How well does the <|sep|> block-marker architecture handle heavily obfuscated HTML or adversarial SEO spam where boilerplate is styled to look like semantic body text?
Have you tested running pulpie-orange-small (210M) quantized (e.g., INT8/FP8) on consumer edge GPUs or CPU-only setups for local RAG pipelines?
Amazing work on open-sourcing the teacher and distilled weights on Hugging Face!
Why not use a plain old html → markdown converter? You can easily strip out ads using CSS /jQuery-like selectors. That would cost zero dollars.
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.
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.
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.
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.
Fixed. Try again. Let me know if any other issues
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.
Funnily enough, that wasn't my first choice either. I A/B tested it with a small group and people understood "up and to the right is better" faster.
Very nice! Thank you for building this.
Leeches.
ongrats on the release! The architectural insight to switch from a bandwidth-bound decoder (token-by-token generation) to a compute-bound encoder (single forward pass over 8k chunks) is brilliant—the 20x speedup and cost drop from $159k to $7.9k per billion pages is massive for web-scale pipelines.
As someone building AI developer utilities and document tools, I have two quick technical questions:
How well does the <|sep|> block-marker architecture handle heavily obfuscated HTML or adversarial SEO spam where boilerplate is styled to look like semantic body text? Have you tested running pulpie-orange-small (210M) quantized (e.g., INT8/FP8) on consumer edge GPUs or CPU-only setups for local RAG pipelines? Amazing work on open-sourcing the teacher and distilled weights on Hugging Face!