The corpus is converted and kept private. The Chadian-language texts must not be published, sold or redistributed until the Alliance Biblique du Tchad signs the written licence. This page shows counts and field layout, not the text.
What the corpus contains
Each language is aligned verse by verse with the Louis Segond 1910 French Bible. French is the pivot, not a source language. The source database holds 18 versions; only these three are in scope, the other Chadian versions belong to other publishers.
Three languages, one French reference
Each row is a verse in the Chadian language and the same verse in French. Sar (mwm) is planned; its text is not in the source database yet.
| ISO 639-3 | Language | Version | Pairs | Train | Val | Test |
|---|
One aligned pair per line
The files are JSONL. Alignment uses the shared Book.Chapter.Verse reference, so any two translations of the same verse form a pair. The two text fields are blanked here; the rest is a real row from the Chadian Arabic to French files.
{
"verse_id": "JHN.3.16",
"book": "John",
"chapter": 3,
"verse": 16,
"source_lang": "shu",
"target_lang": "fra",
"source_text": "…",
"target_text": "…",
"domain": "religion",
"source_version": "شو",
"target_version": "LSG1910"
}
source_version is the publisher's version code (شو, SBA2015, MUAR); target_version is always LSG1910.
What was checked, and what was dropped
| Version | Pairs | Kept | Chapters dropped | Next to a gap | Length outliers |
|---|
A chapter is dropped when its verse numbers differ from the French: one shifted verse would misalign the whole chapter. Verses next to a missing verse are dropped, because a merged verse may hold the missing text. Pairs outside [median ÷ 3, median × 3], or with a text shorter than 4 characters, are dropped. In Chadian Arabic, the source's < > quotation marks are converted to « ».
Read it with the datasets library
The corpus is not published. From a local export, load the JSONL files directly:
from datasets import load_dataset
ds = load_dataset("json", data_files={
"train": "shu-502/train.jsonl",
"validation": "shu-502/validation.jsonl",
"test": "shu-502/test.jsonl",
})
shu = ds["train"]
tchad_parallel_bible.py declares one configuration per language (shu, sba, mua). It reads the folders when they sit next to the script, or from TCHAD_PARALLEL_DATA.
Known limits
- The splits are by whole book, so no verse is in two splits. Test is John, Ruth and Jonah; validation is Romans and Proverbs. Test holds 928 to 1,012 pairs per language: enough for a smoke test, not for a benchmark claim.
- Psalm titles stay in verse 1 where the source puts them, as in the Segond. Chapters whose numbering already differs are dropped. Removing titles would need a per-psalm, per-language table; it has not been done.
- Verse alignment, not sentence alignment: a sentence that crosses a verse boundary is split across rows.
- No linguist has reviewed the pairs. The checks above are automatic.
- Chadian Arabic is written in Arabic script with a local orthography, not Modern Standard Arabic.
- Sar (
mwm) is agreed in principle but absent until its text is obtained.
Rights and licence
The Chadian Arabic, Sar, Ngambay and Moundang Bibles belong to the Alliance Biblique du Tchad. The Alliance has agreed in principle, orally, to this use for these four languages, including commercial licensing. That agreement is being put in writing. Nothing is sold, published or transferred before the written licence is signed; the corpus licence will follow its terms.
The conversion starts from a local database of published versions assembled by the project owner. No rights are claimed from that database or from any Bible app: the written licence is the source of rights. Other versions in the database (Kim, Mbay, Tupuri, Musey and others) belong to other publishers, with whom there is no agreement, and are excluded.
The French Louis Segond 1910 text is public domain. The code (build_from_db.mjs, tchad_parallel_bible.py) and the file layout may be reused freely; the Chadian-language texts may not.
Write to us
We reply in English or French. You can also open your mail app.