pulpie-orange-large

<p align="center"> <img src="https://huggingface.co/feyninc/pulpie-orange-large/resolve/main/pulpie-orange.png" alt="Pulpie Orange" width="360"> </p> <h1 align="center">Pulpie Orange Large</h1> <p align="center"> <b>Pareto-optimal main-content extraction from HTML.</b><br> 2.1B-parameter encoder · 0.873 ROUGE-5 F1 on WebMainBench · the highest-quality Pulpie model (teacher). </p> <p align="center"> <a href="https://github.com/feyninc/pulpie">GitHub</a> · <a href="https://usefeyn.com/blog/pulpie-pareto-optimal-models-for-cleaning-the-web/">Blog</a> · <a href="https://pypi.org/project/pulpie/">PyPI</a> </p>

---

Pulpie Orange Large extracts the main content from raw HTML, stripping navigation, ads, sidebars, and footers. It is an encoder that labels every HTML block as content or boilerplate in a single forward pass, so it approaches state-of-the-art extraction quality while running far faster and cheaper than autoregressive extractors.

At 2.1B parameters it is the strongest single model in the family, scoring 0.873 ROUGE-5 F1 — ahead of Dripper (0.864) at a third the memory-bandwidth cost. It is the teacher from which Orange Base and Orange Small are distilled. Use it when you want maximum quality; for most production workloads Orange Small matches it within ~1 F1 point at a fraction of the cost.

Usage

The easiest way to use this model is through the pulpie package:

pip install pulpie

from pulpie import Extractor

extractor = Extractor(model="orange-large")
result = extractor.extract(html)

print(result.markdown)                # clean Markdown
print(result.html)                    # clean HTML
print(result.n_main, result.n_other)  # blocks kept vs dropped

Extractor auto-detects CUDA, Apple MPS, then CPU. See the GitHub README for batch and multi-GPU usage.

How it works

Pulpie runs a four-stage pipeline:

1. Simplify — remove scripts, styles, and formatting noise; tag each block with a unique ID.
2. Chunk — pack blocks into sequences of up to 8,192 tokens separated by <|sep|> markers (~80% of pages fit in one chunk).
3. Classify — a single encoder forward pass labels every block (at its <|sep|> position) as content or boilerplate.
4. Reconstruct — return the kept blocks as HTML, or convert to Markdown.

This model is a token-classification head over EuroBERT-2.1B, fine-tuned on 14,959 Common Crawl pages with block-level labels (DeepSeek V3.2 labeled, Dripper cross-validated), using class-weighted cross-entropy.

Benchmarks

WebMainBench, English subset (6,647 pages), ROUGE-5 F1:

Full analysis in the blog post.

Model family

Acknowledgements

Pulpie builds directly on the work of the MinerU-HTML and Dripper team (Ma et al., 2025). Their simplify_html preprocessing, block-level annotation scheme, and the WebMainBench benchmark are foundational to this work. Built on EuroBERT (Boizard et al., 2025).

Citation

@note{pulpie2026,
  title  = {Pulpie: Pareto-Optimal Models for Cleaning the Web},
  author = {Minhas, Bhavnick and Nigam, Shreyash and Feyn Research},
  year   = {2026},
  venue  = {Feyn Field Notes}
}

Built by Feyn. Model weights and the pulpie library are licensed under Apache 2.0.