README.md

---
language:
  - it
license: cc-by-4.0
task_categories:
  - text-generation
tags:
  - italian
  - modotai
  - pretraining
  - corpus
size_categories:
  - 1M<n<10M
---

# Italian-Common-Corpus

The Italian dataset with the highest density of useful information per token. Built by [ModotAI](https://things-ai.org) for training Italian language models.

## Subsets

| Subset | File | Documents | Words | Description |
|--------|------|-----------|-------|-------------|
| **Web Crawl** | `icc-web.parquet` | ~27K | ~17M | Italian sources: news, tech, science, culture, law, food, sport |
| **Wikipedia IT** | `wiki-it-clean.parquet` | ~1.35M | ~698M | Cleaned Italian Wikipedia — removed Notes, Bibliography, Voci correlate, stub articles |

**Total: 1,377,369 documents, 715,313,213 words (~929M tokens)**

## Usage

```python
from datasets import load_dataset

# Load web crawl subset
icc = load_dataset("ThingsAI/Italian-Common-Corpus", data_files="data/icc-web.parquet", split="train")

# Load Wikipedia subset
wiki = load_dataset("ThingsAI/Italian-Common-Corpus", data_files="data/wiki-it-clean.parquet", split="train")

# Load everything
full = load_dataset("ThingsAI/Italian-Common-Corpus", split="train")

print(icc[0])
```


## Quality Pipeline

1. **Async crawling** — 20 parallel workers, rate-limited per domain
2. **Text extraction** — Trafilatura with precision mode
3. **Spam detection** — 13 regex patterns + statistical analysis
4. **Quality scoring** — Lexical richness, sentence structure, punctuation (0-100)
5. **Deduplication** — MD5 hash + partial content matching
6. **Wikipedia cleaning** — Section removal (Notes, Bibliography, See also, External links)

Only documents scoring ≥50/100 are included.

## License

CC-BY-4.0

## Citation

```bibtex
@dataset{modotai_icc_2026,
  author = {ModotAI},
  title = {Italian-Common-Corpus},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/datasets/ThingsAI/Italian-Common-Corpus}
}
```