Request access

Describe your intended use below. Downloads unlock after approval.

Weights are released under CC BY-NC-SA 4.0 for non-commercial research and education. Credit NileAGI, keep adaptations under the same licence, and contact us before any commercial product or paid service. Do not use the model to harm people, break the law, or present outputs as human writing.

Sukuma MT Lite

Authors: Zephania and Isack Odero

Compact Swahili → Sukuma translation for the macron orthography (ā ē ī ō ū). Same one-way task as the quality SKU, with a smaller footprint and lower chrF2.

Prefer nileagi/nileagi-suk-mt when quality matters more than size. For English, pivot English → Swahili first.


Quality sibling	`nileagi/nileagi-suk-mt`
Collection	nileagi/nileagi-suk

Summary


Task	Machine translation
Direction	Swahili → Sukuma
Language codes	`swh_Latn` → `suk_Latn`
Pairs	31,102 · train 29,962 · val 993 · test 147
Split rule	whole document groups
Orthography	Latin with vowel macrons
Primary metric	chrF2

Evaluation

Same held-out document-group test (147 pairs). Quote test chrF2.

Metric	Lite (this repo)	Quality `nileagi-suk-mt`
Test chrF2	34.8	42.6
Test BLEU	12.7	20.0
Test eval loss	2.46	1.97
Val chrF2 (993)	39.1	46.3
Val BLEU (993)	16.3	23.2

Use this SKU for a lighter footprint, not for the best text.

How to use

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("nileagi/nileagi-suk-mt-lite")
model = AutoModelForSeq2SeqLM.from_pretrained("nileagi/nileagi-suk-mt-lite")

tok.src_lang = "swh_Latn"
inputs = tok("Habari yako?", return_tensors="pt")
bos = tok.convert_tokens_to_ids("suk_Latn")
out = model.generate(
    **inputs,
    forced_bos_token_id=bos,
    max_new_tokens=128,
)
print(tok.decode(out[0], skip_special_tokens=True))

Set swh_Latn and force suk_Latn.

Intended use

Constrained Swahili → Sukuma where the quality SKU is too heavy

Out of scope

Anything the quality SKU cannot do: Sukuma → Swahili, direct English → Sukuma, chat, legal/medical, conversational Sukuma.

Limitations

Same family as the quality SKU: macron orthography, one-way, extra spaces inside words, small document-group test. Short lines are often looser than on the quality model.

Access

Weights are gated under CC BY-NC-SA 4.0.

Open Request access at the top of this page and describe your intended use.
After approval, authenticate (huggingface-cli login or HF_TOKEN) and download or load the model.

Attribute NileAGI when you use or redistribute the weights. Adapted weights must stay under CC BY-NC-SA 4.0. Commercial licensing: hi@nileagi.com.

License

Weights: CC BY-NC-SA 4.0.
See LICENSE and NOTICE.md.

Citation

@misc{nileagi-suk-mt-lite-2026,
  title        = {Sukuma machine translation lite (Swahili to Sukuma)},
  author       = {Zephania and Isack Odero},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/nileagi/nileagi-suk-mt-lite},
  note         = {NileAGI}
}

Downloads last month: 29

Safetensors

Model size

0.6B params

Tensor type

F32

Collection including nileagi/nileagi-suk-mt-lite

nileagi-suk

Collection

NileAGI Sukuma dataset and small STT, TTS, and MT models. • 4 items • Updated 2 days ago

nileagi
/

nileagi-suk-mt-lite