Update README.md

README.md

@@ -16,22 +16,13 @@ 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.
+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.
 
-|             |       |
-|---|---|
-| **Train rows**   | 2,482,568 |
-| **Val rows**     | 272,875 |
-| **Languages**    | 334 (ISO 639-3) |
-| **Format**       | Apache Parquet (zstd-compressed) |
-| **License**      | Per-source (see attribution table) |
-
-## Files
-
-| File | Rows | Size | Description |
-|---|---:|---:|---|
-| `train.parquet` | 2,482,568 | 1.08 GB | Provenance schema (10 columns) |
-| `val.parquet`   |   272,875 | 119 MB  | Same schema |
+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
 
@@ -48,52 +39,16 @@ The training data for [PleIAs/CommonLingua](https://huggingface.co/PleIAs/Common
 | `creator`      | string | Author / organisation, nullable |
 | `date`         | string | Publication / extraction date, nullable |
 
-## Source attribution
-
-Wikipedia provides the bulk (≈ 94 %); the long tail is filled by [Common Corpus](https://huggingface.co/datasets/PleIAs/common_corpus) collections and a handful of dedicated minority-language corpora.
-
-| Source | Rows | License | Notes |
-|---|---:|---|---|
-| Wikipedia              | 2,323,301 | CC-BY-SA 4.0 | Per-article paragraphs across 300+ language editions |
-| OpenAlex               |    29,733 | per-journal  | Mostly CC-BY academic content (Indonesian/Malaysian, African, English) |
-| Pralekha               |    27,076 | CC-BY-SA 4.0 | Indic-language paragraphs |
-| VOA Africa             |    24,341 | Public Domain (US Federal) | Voice of America Africa-language broadcasts |
-| Cultural Heritage      |    16,381 | per-collection | Digitised heritage archives (Tocharian, Larth Etruscan, ePSD2/CDLI, Thesaurus Linguae Aegyptiae) |
-| Perseus                |    11,638 | CC-BY-SA 4.0 | Classical languages (Latin, Ancient Greek, Old Church Slavonic, Gothic) |
-| OpenPecha              |     8,317 | Public Domain / CC | Tibetan and Sanskrit (incl. Ambuda) |
-| WaxalNLP               |     5,373 | CC-BY-SA 4.0 | West African ASR transcripts |
-| eBible                 |     4,877 | Public Domain | Bible translations, public-domain editions only |
-| Chinese-Court-Decisions|     4,474 | Public Domain | Chinese court rulings |
-| StackExchange          |     3,312 | CC-BY-SA 4.0 | Q&A across StackExchange sites |
-| Project Ben-Yehuda     |     2,701 | Public Domain | Hebrew literature |
-| USPTO                  |     2,029 | Public Domain | US patent filings |
-| US-PD-Newspapers       |     1,834 | Public Domain | US public-domain newspapers |
-| HAL                    |     1,575 | per-license | French academic abstracts |
-| French-PD-Newspapers   |     1,539 | Public Domain | |
-| English-PD             |     1,353 | Public Domain | |
-| dotgov                 |     1,130 | Public Domain | US .gov web text |
-| French-PD-diverse      |     1,065 | Public Domain | |
-| Sefaria                |       947 | Public Domain | Hebrew religious corpus |
-| Court Listener         |       931 | Public Domain | US court filings |
-| Krike-Krake            |       744 | Public Domain | Tachelhit / Berber documentation |
-| Deutsches Zeitungsportal |     615 | Public Domain | German newspaper archive |
-| govinfo                |       572 | Public Domain | US Government publications |
-| Creative Commons Common Crawl (CCC) | 561 | per-license CC | |
-| enevaeldens_nyheder    |       496 | Public Domain | Danish historical news |
-| German-PD              |       485 | Public Domain | |
-| French-PD-Books        |       368 | Public Domain | |
-| French Open Data       |       339 | per-license | |
-| Spanish-Science-Pile   |       323 | per-journal | |
+## Composition
 
-The remaining ~90 sources cover the long tail of [Common Corpus](https://huggingface.co/datasets/PleIAs/common_corpus) collections (Italian-PD, Latin-PD, Spanish-PD-*, Open Korean Historical Corpus, Wikisource, UN-Digital-Library, …) and account for ~3,000 rows total. Per-row licensing is preserved in the `license` and `open_type` columns of the parquet itself.
+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)
 
-## Reproducing the model
+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"])
-# train ByteHybrid using github.com/PleIAs/bytehybrid-lid/train_lid.py
 ```
 
 ## License & responsible use