Datasets:

PleIAs
/

CommonLingua-Train

+---
+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
+The training data for [PleIAs/CommonLingua](https://huggingface.co/PleIAs/CommonLingua) — a byte-level language identification model for 334 languages. **2.48 M paragraphs**, sourced exclusively from open-licensed and public-domain corpora.
+|             |       |
+|---|---|
+| **Train rows**   | 2,482,568 |
+| **Val rows**     | ~270 k |
+| **Languages**    | 334 (ISO 639-3) |
+| **Sources**      | 30 distinct |
+| **Format**       | Apache Parquet |
+| **License**      | Per-source (see attribution table below) |
+## Files
+| File | Rows | Size | Description |
+|---|---:|---:|---|
+| `train.parquet`           | 2,482,568 | 1.31 GB | Full provenance schema (12 columns) |
+| `val.parquet`             |  ~270 k   | 144 MB  | Validation split, same schema |
+| `train_text_lang.parquet` | 2,482,568 | 1.14 GB | Training-time projection (`text`, `lang`) |
+| `val_text_lang.parquet`   |  ~270 k   | 125 MB  | Same projection for val |
+The provenance files are the **canonical artefact**. The two-column projections are convenience copies for users who only want to retrain a classifier and don't care about source / license / metadata; they were derived from the provenance files at the moment training launched.
+## Schema (provenance files)
+| 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) |
+| `language_method`  | string | How `lang` was assigned: `metadata` / `LLM` / `consensus` / `provenance` / `journal_provenance` / `manual_latex` |
+| `source`           | string | Top-level source identifier (e.g. `Wikipedia`, `Pralekha`, `VOA_Africa`) |
+| `identifier`       | string | Item-level identifier within the source (URL / DOI / shelfmark / file id), nullable |
+| `title`            | string | Item title, nullable |
+| `collection`       | string | Sub-collection within the source (e.g. `Voice of America`, `Comprehensive Edition of Tocharian Manuscripts`) |
+| `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 |
+| `section_name`     | string | Sub-section within the document (e.g. Wikipedia article section), nullable |
+## `language_method` values
+| Method | Rows | Meaning |
+|---|---:|---|
+| `metadata`           | 2,184,648 | Label inherited from the source's collection-level metadata (e.g. Wikipedia language code) |
+| `LLM`                |   183,071 | Label assigned (or verified) by Gemini 2.5 Flash row-by-row |
+| `consensus`          |    67,624 | Label confirmed by agreement between multiple LID models on a Wikipedia resample |
+| `provenance`         |    26,725 | Label deduced from per-row provenance metadata in the original source |
+| `journal_provenance` |    20,000 | Label deduced from journal-name provenance (Indonesian / Malaysian academic journals) |
+| `manual_latex`       |       500 | Manually labelled `__NOT_LANGUAGE__` / LaTeX-source rows |
+## Source attribution
+| Source | Rows | License | Notes |
+|---|---:|---|---|
+| Wikipedia (raw)                  | 2,093,518 | CC-BY-SA 4.0 | Per-article text, sampled across language editions |
+| Wikipedia_Gemini_verified        |    93,988 | CC-BY-SA 4.0 | Targeted resample with Gemini per-row verification of contaminated classes |
+| Wikipedia_resample_Gemini        |    67,731 | CC-BY-SA 4.0 | Class-balanced resample, Gemini-verified |
+| Wikipedia_resample               |    67,624 | CC-BY-SA 4.0 | Class-balanced resample (consensus method) |
+| Pralekha                         |    27,042 | CC-BY-SA 4.0 | Indic-language paragraphs |
+| VOA Africa                       |    24,341 | Public Domain (US Federal) | Voice of America Africa-language broadcasts |
+| Common Corpus 1 v3 adversarial   |    17,953 | per-source (CC) | Hand-picked adversarial paragraphs from Common Corpus |
+| Cultural Heritage                |    16,381 | Public Domain / per-collection | Digitised cultural-heritage archives |
+| Perseus PROIEL                   |    11,638 | CC-BY-SA 4.0 | Classical languages (Latin, Ancient Greek, Old Church Slavonic, Gothic, …) |
+| OpenAlex (Malaysian journals)    |    10,000 | per-journal (mostly CC-BY) | Indonesian/Malay disambiguation data, journal-name provenance |
+| OpenAlex (Indonesian journals)   |    10,000 | per-journal (mostly CC-BY) | Companion to the above |
+| OpenPecha / Ambuda               |     8,317 | Public Domain / CC | Tibetan, Sanskrit |
+| Common Corpus 1 Haiku v5         |     7,497 | per-source (CC) | Haiku-validated minority-language paragraphs |
+| OpenAlex (African academic)      |     6,193 | per-journal | African academic publishing, English-heavy filtered out |
+| WaxalNLP ASR                     |     5,373 | CC-BY-SA 4.0 | West African ASR transcripts (residual after extensive cleanup — see History below) |
+| eBible (Public Domain)           |     4,877 | Public Domain | Bible translations, public-domain editions only |
+| OpenAlex English                 |     2,705 | per-journal | English baseline from academic publishing |
+| Project Ben-Yehuda (Hebrew)      |     2,701 | Public Domain | Hebrew literature corpus |
+| French Research HAL              |     1,575 | per-license | French academic abstracts |
+| Sefaria Tanakh                   |       947 | Public Domain | Hebrew religious corpus |
+| Krike-Krake (Barat 1977)         |       744 | Public Domain | Tachelhit/Berber documentation |
+| OpenAlex LaTeX                   |       500 | n/a | Manually labelled `__NOT_LANGUAGE__` (LaTeX source) |
+| US-PD-Newspapers                 |       348 | Public Domain | Long-tail language samples |
+| Common Corpus Adversarial        |       243 | per-source | |
+| Geste                            |       202 | Public Domain | Old French |
+| English-PD                       |       108 | Public Domain | |
+| German-PD                        |         8 | Public Domain | |
+| French-PD-Newspapers             |         6 | Public Domain | |
+| French-PD-Books                  |         6 | Public Domain | |
+| Italian-PD                       |         2 | Public Domain | |
+The single largest source is Wikipedia (≈ 93 % of rows when raw + resample variants are summed). The long tail (~7 %) is what makes coverage of low-resource and ancient languages possible.
+## Reproducing the model
+```python
+import pyarrow.parquet as pq
+train = pq.read_table("train.parquet", columns=["text", "lang"])
+val   = pq.read_table("val.parquet",   columns=["text", "lang"])
+# train ByteHybrid using github.com/PleIAs/bytehybrid-lid/train_lid.py
+```
+## License & responsible use
+CommonLingua-Train aggregates 30 sources whose licenses range from **public domain** (most newspaper / heritage / Bible / classical-philology corpora) to **CC-BY-SA 4.0** (Wikipedia and most aggregated corpora). Some OpenAlex rows inherit the journal's per-publication license — most are CC-BY but please consult the `license` and `open_type` columns of each row before redistributing.
+If you redistribute derived versions, please keep the per-row `license`, `creator`, and `identifier` columns intact for downstream attribution.
+## History
+This dataset is the v7.2.1 iteration. Earlier iterations (v6.3.x) carried collection-level labels at face value, which led to mislabelled African-language data (Wax̌alNLP collections under `kik`/`mas`/`sog`/etc.). The v7 line dropped those classes outright and re-validated the rest with Gemini 2.5 Flash. The current dataset retains only ~5 k rows from the original WaxalNLP collection, after that cleanup.
+## 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}
+}
+```