2ADT-Consulting/susu-parallel

Susu (Soussou) Parallel and Monolingual Corpus

A multi-source corpus for Susu (Soussou; ISO 639-3 sus), a Mande language of Guinea that is absent from NLLB-200 and from commercial MT systems. Built to train 2ADT-Consulting/nllb-susu-v2, one of the first open neural MT systems for Susu.

Configurations

Config	Split	#rows	Columns
sus-fr	train / validation / test	114,503 / 1,000 / 1,000	sus, fr
sus-en	train / validation / test	111,013 / 991 / 992	sus, en
monolingual	train	28,724	sus
lexicon	train	~7,000	sus_officiel, sus_ortho_fr, fr, ...

from datasets import load_dataset
ds = load_dataset("2ADT-Consulting/susu-parallel", "sus-fr")
print(ds["test"][0])   # {'sus': '...', 'fr': '...'}

Sources and construction

• Bible translations (Susu ↔ fr/en) aligned at verse level via verse keys, retrieved using the Africa Corpus Builder tool (AfriSpeech), which sources public Bible translations from YouVersion. This is the backbone of the parallel data.
• Bilingual dictionary (~7,000 Susu–French entries), including both the official orthography and a French-style spelling (sus_ortho_fr).
• Everyday dialogues extracted from a Peace-Corps Susu language-course manual, plus additional Christian-education booklets.
• Orthography-aware augmentation: informal "French-style" (SMS) variants of Susu sentences generated with deterministic rules (x→kh, u→ou, ɛ→è, ɔ→o, ɲ→gn, ...), so models learn to read informal input while producing the official orthography.

Pairs with an out-of-range source/target length ratio are filtered (211 removed for sus–fr, 522 for sus–en). Vocabulary: 8,690 unique word types, 757,387 tokens.

Intended use and limitations

• Intended for machine translation and low-resource NLP research on Susu.
• Register bias: the parallel data is dominated by biblical text; models and automatic metrics computed against these references over-reward the religious register. The everyday/dictionary/dialogue data partially mitigates this.
• No native-speaker validation yet. Automatic references may contain errors.

License and citation

This dataset's text is not covered by NLLB-200's CC-BY-NC-4.0 license — that license applies to the model weights we separately publish (nllb-susu-v2), not to this text. The Bible-derived portion originates from YouVersion and remains subject to YouVersion's terms of service, independently of any model license; we retrieved it via the Africa Corpus Builder tool and did not access YouVersion directly. The dictionary and course-manual portions carry their own source terms. Review all applicable terms before redistributing derived data.

If you use the Bible-derived portion, please also credit the tool used to retrieve it:

@misc{africacorpusbuilder,
  title  = {{Africa Corpus Builder}: verse-aligned text for 693 African languages},
  author = {{AfriSpeech}},
  note   = {\url{https://github.com/AfriSpeech/africa-corpus-builder}},
  year   = {2026}
}

To cite this corpus itself:

@misc{susu-corpus,
  title  = {Susu (Soussou) Parallel and Monolingual Corpus},
  author = {TODO},
  year   = {2026},
  note   = {\url{https://huggingface.co/datasets/2ADT-Consulting/susu-parallel}}
}