README.md

---
license: apache-2.0
library_name: transformers
pipeline_tag: token-classification
tags:
  - html
  - content-extraction
  - web-scraping
  - boilerplate-removal
  - encoder
  - eurobert
  - rag
base_model: EuroBERT/EuroBERT-610m
language:
  - en
metrics:
  - rouge
---

<p align="center">
  <img src="https://huggingface.co/feyninc/pulpie-orange-base/resolve/main/pulpie-orange.png" alt="Pulpie Orange" width="360">
</p>

<h1 align="center">Pulpie Orange Base</h1>

<p align="center">
  <b>Pareto-optimal main-content extraction from HTML.</b><br>
  610M-parameter encoder · 0.863 ROUGE-5 F1 on WebMainBench · the balanced Pulpie model.
</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 Base** 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 610M parameters it sits between [Orange Small](https://huggingface.co/feyninc/pulpie-orange-small) and [Orange Large](https://huggingface.co/feyninc/pulpie-orange-large), scoring 0.863 ROUGE-5 F1. For most use cases Orange Small offers a better speed/quality trade-off; choose Base when you want a little more headroom than Small without the cost of the 2.1B teacher.

## Usage

The easiest way to use this model is through the [`pulpie`](https://pypi.org/project/pulpie/) package:

```bash
pip install pulpie
```

```python
from pulpie import Extractor

extractor = Extractor(model="orange-base")
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](https://github.com/feyninc/pulpie) 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-610m](https://huggingface.co/EuroBERT/EuroBERT-610m), distilled from the 2.1B [Pulpie Orange Large](https://huggingface.co/feyninc/pulpie-orange-large) teacher (KL-divergence 0.7 + hard-label cross-entropy 0.3, temperature 2.0).

## Benchmarks

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

| Model | Params | ROUGE-5 F1 | Throughput (L4) |
|-------|--------|------------|-----------------|
| [Pulpie Orange Large](https://huggingface.co/feyninc/pulpie-orange-large) | 2.1B | 0.873 | 1.3 pages/sec |
| Dripper | 0.6B | 0.864 | 0.68 pages/sec |
| **Pulpie Orange Base (this model)** | **610M** | **0.863** | **3.9 pages/sec** |
| [Pulpie Orange Small](https://huggingface.co/feyninc/pulpie-orange-small) | 210M | 0.862 | 13.7 pages/sec |
| magic-html | - | 0.700 | - |
| Trafilatura | - | 0.619 | - |

Full analysis in the [blog post](https://usefeyn.com/blog/pulpie-pareto-optimal-models-for-cleaning-the-web/).

## Model family

| Model | Params | ROUGE-5 F1 | Use case |
|-------|--------|------------|----------|
| [pulpie-orange-small](https://huggingface.co/feyninc/pulpie-orange-small) | 210M | 0.862 | **Recommended** — best value, fastest |
| [pulpie-orange-base](https://huggingface.co/feyninc/pulpie-orange-base) | 610M | 0.863 | Balanced |
| [pulpie-orange-large](https://huggingface.co/feyninc/pulpie-orange-large) | 2.1B | 0.873 | Highest quality (teacher) |

## 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](https://huggingface.co/EuroBERT) (Boizard et al., 2025).

## Citation

```bibtex
@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](https://usefeyn.com). Model weights and the [pulpie](https://pypi.org/project/pulpie/) library are licensed under Apache 2.0.