pulpie 0.0.2

First release under the feyninc org.

• Ships pulpie.markdown (the [markdown] extra now matches the docs).
• Extractor/Pipeline DX fixes: markdown cleanup, error propagation, extract_batch, dataclass results.
• MinerU-HTML-parity simplify/reconstruct port (byte-parity tested against the vendored oracle).
• Apple MPS auto-detection.
• Model IDs now resolve to feyninc/pulpie-orange-{small,base,large}.

Install: pip install "pulpie[markdown]"

pulpie v0.0.1

First public release of pulpie — fast, CPU-friendly web content extraction from HTML using encoder models.

Install

pip install pulpie            # core
pip install pulpie[markdown]  # + markdown output

What it does

pulpie extracts the main content from raw HTML and returns clean HTML or markdown. It classifies HTML blocks with small encoder models (the Orange family), reaching autoregressive-extractor quality at a fraction of the cost.

from pulpie import Extractor

extractor = Extractor()             # downloads pulpie-orange-small (210M) on first use
result = extractor.extract(html)
print(result.markdown)              # clean markdown
print(result.html)                  # clean HTML
print(result.n_main, result.n_other)

MinerU-HTML parity

pulpie's simplify/reconstruct faithfully reproduce MinerU-HTML's simplify_html / map_to_main (Apache-2.0), the segmentation format the Orange models were distilled on. Verified two ways:

• Offline byte-parity: 65/65 tests reproduce MinerU's exact output across 13 real-page fixtures.

• End-to-end ROUGE-5 on full WebMainBench English (6,647 pages, orange-small):

  Simplification	ROUGE-5 F1
  MinerU-HTML (reference)	0.8625
  pulpie (this release)	0.8625

  Byte-faithful end-to-end (Δ vs MinerU = +0.0000), a +0.14 improvement over the prior simplification.

Model zoo

Name	HF repo	Size	Notes
Orange Small	chonkie-ai/pulpie-orange-small	210M	Best value; default
Orange Base	chonkie-ai/pulpie-orange-base	610M	Distilled
Orange Large	chonkie-ai/pulpie-orange-large	2.1B	Highest quality

Credits

Built on MinerU-HTML / Dripper (Ma et al., 2025) — their simplify_html preprocessing, block-annotation scheme, and the WebMainBench benchmark are foundational. Thank you for releasing your tools and data.