alaminerca/sango-french-bible-parallel

SFPC: Sango-French Parallel Corpus

The first quality-filtered, verse-aligned Sango-French parallel corpus, constructed for neural machine translation research. This dataset directly addresses the "Sango Problem" identified by Meta's NLLB-200 project — the failure of cross-lingual transfer for a linguistically isolated Creole language.

Associated resources:

• —Model: alaminerca/nllb-sango-french
• —Demo: Sango-French Translator
• —Paper: SangoNMT: Parameter-Efficient Domain Adaptation of NLLB-200 for Sango, an Isolated Creole Language (Submitted, 2026)

Dataset Summary

Splits

Splits are at book level (not random) to prevent data leakage. Adjacent Bible verses share vocabulary and themes; random splitting would inflate evaluation scores. Validation and test books cover all five major Biblical genres.

Data Format

Each example is a JSON object with the following fields:

{
  "id": "genesis_1_1_fr2sg",
  "source_lang": "fra_Latn",
  "target_lang": "sag_Latn",
  "source_text": "Au commencement Dieu créa les cieux et la terre.",
  "target_text": "Na tongo nda ni, Nzapa asara yayu na sese.",
  "book": "Genesis",
  "chapter": 1,
  "verse": 1
}

The dataset is bidirectional: each verse pair appears twice (French→Sango and Sango→French), enabling a single model to translate in both directions.

Construction Pipeline

The corpus was constructed through a four-stage pipeline:

1. 1.PDF Extraction — Text extracted from Sango and French Bible PDFs using PyMuPDF with regex-based verse identification
2. 2.Verse-Level Alignment — Verses aligned by standardized book-chapter-verse addresses across both Bibles
3. 3.Encoding Normalization — Systematic UTF-8 mojibake patterns corrected (40+ known corruption patterns)
4. 4.Quality Filtering — Three filters applied: minimum 10 characters per side, length ratio within 1:3 to 3:1, content presence in both languages. Removed 849 pairs (3.9%)

Alignment Statistics

Genre Distribution

Counts are bidirectional examples.

Text Length Statistics

Usage

from datasets import load_dataset

dataset = load_dataset("alaminerca/sango-french-bible-parallel")

# Access splits
train = dataset["train"]
val = dataset["validation"]
test = dataset["test"]

# Example
print(train[0])

Use with NLLB-200

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("alaminerca/nllb-sango-french")
tokenizer = AutoTokenizer.from_pretrained("alaminerca/nllb-sango-french")

# Get a test example
example = dataset["test"][0]
tokenizer.src_lang = example["source_lang"]

inputs = tokenizer(example["source_text"], return_tensors="pt", max_length=256, truncation=True)
tgt_id = tokenizer.convert_tokens_to_ids(example["target_lang"])
output = model.generate(**inputs, forced_bos_token_id=tgt_id, max_new_tokens=256, num_beams=3)
print(tokenizer.decode(output[0], skip_special_tokens=True))

About Sango

Sango (ISO 639-3: sag) is the national language of the Central African Republic, spoken by approximately 5.5 million people. It is a creole-like language derived from Ngbandi (Ubangian family), characterized by tonal distinctions, minimal morphology, and SVO word order.

The NLLB-200 project identified Sango as the only Creole language for which no similar high-resource language could be found for cross-lingual transfer, resulting in the lowest Creole performance (5.3 BLEU with 1.9M mined sentences). This dataset provides high-quality, human-translated parallel data as an alternative to noisy mined data.

Limitations

• —Domain: Biblical text only. The formal register of Biblical Sango differs from everyday conversational usage.
• —Register: Written Sango is almost exclusively formal; this corpus reflects that bias.
• —Tone: Sango is tonal but tone markings are inconsistent in the source texts.
• —Coverage: 66 of 73 standard Biblical books (some books had 0% alignment rate).

Citation

@article{mouhamad2026sangonmt,
  title={SangoNMT: Parameter-Efficient Domain Adaptation of NLLB-200 for Sango, an Isolated Creole Language},
  author={Mouhamad, Alim Al-Amine and Alkhodre, Ahmad B. and Alsaawy, Yazed},
  year={2026},
  note={Submitted}
}

Authors

Alim Al-Amine Mouhamad, Ahmad B. Alkhodre, Yazed Alsaawy — Department of Computer Science, Islamic University of Madinah, KSA