README.md

---
dataset_info:
  features:
  - name: text
    dtype: string
  - name: id
    dtype: string
  - name: url
    dtype: string
  - name: edu_score
    dtype: float32
  - name: stem_score
    dtype: float32
  - name: toxic_score
    dtype: float32
  splits:
  - name: train
    num_bytes: 285576324067
    num_examples: 96975210
  download_size: 164777356452
  dataset_size: 285576324067
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# 📚 ClassiCC-PT: Classified Common Crawl Corpus for Portuguese

## 📖 Overview

ClassiCC-PT (Classified Common Crawl – Portuguese) is a large-scale web corpus containing ~120B Portuguese tokens extracted from Common Crawl snapshots. It is specifically curated for training large language models in Portuguese, with a focus on data quality, language specificity, and targeted filtering.

This corpus was created as part of a study on continued pretraining for adapting English-trained LLMs to Portuguese.


## 🏗 Dataset Construction

Source Snapshots: CC-2021-31, CC-2021-39, CC-2022-40
Steps:

- **Language Filtering**

      Selected only pages tagged with Portuguese in Common Crawl metadata (~2% of each CC crawl).

- **HTML to Text Extraction**

      Used Trafilatura to remove boilerplate and extract main content.

- **Deduplication**

      Applied MinHash intra-crawl deduplication (removing ~40% duplicates).

- **Neural-Based Filtering**

      Developed three BERTimbau-based classifiers for:

      Educational content (ClassiCC-PT-edu)

      STEM content (ClassiCC-PT-STEM)

      Toxic content (ClassiCC-PT-toxic)

      Classifiers were  trained on GPT-4o-annotated Portuguese data.

**Final Corpus**

Retained ~106M documents / ~125B tokens ( Llama 2 tokenizer)


## 🚀 Performance Impact

When used for continued pretraining of TinyLlama-1.1B (1T EN tokens), ClassiCC-PT improved Portuguese benchmark performance (Poeta v1) significantly, outperforming mC4-PT and matching ClueWeb-22-PT. The model trained with ClassiCC-PT is called Curió 1.1B and is available at huggingface.

| Model | Training Regimen | Poeta v1 NPM |
| ---------------------------- | ----------------- | ------------ |
| TinyLlama-1T (EN) | – | 17.4 |
| mC4-PT | cont. pretraining | \~20 |
| ClueWeb-22-PT | cont. pretraining | \~27 |
| **ClassiCC-PT** (Curió-1.1B) | cont. pretraining | **27.1** |



## 📥 Download & Usage

```
from datasets import load_dataset

ds = load_dataset("ClassiCC-Corpus/ClassiCC-PT", split="train")

print(ds[0])

# {
# 'text': '...',
# 'id': '...',
# 'url': '...',
# 'edu_score': 4.0,
# 'stem_score': 1.0,
# 'toxic_score': 0.0
# }

```

## 📜 Citation

If you use ClassiCC-PT, please cite:
```
@article{almeida2025building,
  title={Building High-Quality Datasets for Portuguese LLMs: From Common Crawl Snapshots to Industrial-Grade Corpora},
  author={Almeida, Thales Sales and Nogueira, Rodrigo and Pedrini, Helio},
  journal={Journal of the Brazilian Computer Society},
  volume={31},
  number={1},
  pages={1246--1262},
  year={2025}
}
```

## Acknowledgements

We thank the google TRC program, which generously granted us the necessary resources for the development of this research.