---
license: other
license_name: per-source
language:
- multilingual
size_categories:
- 1M<n<10M
task_categories:
- text-classification
tags:
- language-identification
- common-corpus
- african-languages
pretty_name: CommonLingua-Train
---

# CommonLingua-Train

This is the training dataset for [PleIAs/CommonLingua](https://huggingface.co/PleIAs/CommonLingua) — a byte-level language identification model for 334 languages. It is composed of 2.48 M paragraphs, sourced exclusively from Wikipedia and other open-licensed and public-domain corpora extracted from Common Corpus.

The training dataset was developed iteratively from the initial Structured Wikipedia data subset. Some of the decisions that account for SOTA-level performance includes:
* Filtering of Wikipedia sources, including widespread generated content in some versions and multilingual contaminations.
* Extension to non-encyclopedic sources and formats, especially long documents with OCR errors from Common corpus.
* Additions of low resource language resources, especially coming from Africa, re-identified in Common Corpus.
* Targeted sampling of frequent language confusions, especially between Indonesian and Malay thanks so scientific papers.

## Schema

| Column | Type | Description |
|---|---|---|
| `text`         | string | Paragraph text, raw UTF-8 (truncated to 512 bytes by the trainer) |
| `lang`         | string | ISO 639-3 label (the training target) |
| `source`       | string | Top-level source name (e.g. `Wikipedia`, `Pralekha`, `OpenAlex`) |
| `identifier`   | string | Item-level identifier (URL / DOI / shelfmark / file id), nullable |
| `title`        | string | Item title, nullable |
| `collection`   | string | Sub-collection within the source, nullable |
| `license`      | string | Per-row license string |
| `open_type`    | string | `public_domain`, `cc_by`, `cc_by_sa`, etc. |
| `creator`      | string | Author / organisation, nullable |
| `date`         | string | Publication / extraction date, nullable |

## Composition

The core dataset is from Wikipedia (2,323,301). Additional major inclusions include OpenAlex (30,000 samples, mostly CC-BY academic content in Indonesian/Malaysian and African languages) and some new multilingual subsets we added for the Global Common Corpus update (VOA Africa, Pralekha)

Training can be reproduced by keeping only the text and lang columns:

```python
import pyarrow.parquet as pq
train = pq.read_table("train.parquet", columns=["text", "lang"])
val   = pq.read_table("val.parquet",   columns=["text", "lang"])
```

## License & responsible use

CommonLingua-Train aggregates open-licensed and public-domain corpora. Most sources are either CC-BY-SA 4.0 (Wikipedia, Pralekha, Perseus, WaxalNLP, StackExchange) or public domain (newspapers, patents, government publications, religious texts, classical philology). Some OpenAlex rows inherit the journal's per-publication license — most are CC-BY.

If you redistribute derived versions, please keep the per-row `license`, `creator`, and `identifier` columns intact for downstream attribution.

## Citation

```bibtex
@misc{commonlingua-train,
  author = {{PleIAs}},
  title  = {CommonLingua-Train: A multi-source open dataset for byte-level language identification},
  year   = {2026},
  url    = {https://huggingface.co/datasets/PleIAs/CommonLingua-Train}
}
```