Runyambo Beside Swahili: Lexical Translation Both Ways

•Zephania Reuben and Isack Odero

Runyambo (Orunyambo) is a Bantu language of north-western Tanzania. Speakers already move through Swahili as the regional written bridge—so Swahili–Runyambo bitext is the practical door into a translation stack. Public parallel sentence data for this pair is scarce. Today we open what we can build honestly: a research preview of lexical and short-phrase Swahili ↔ Runyambo translation, not a claim of full-sentence production MT.

Three open checkpoints go out together on Hugging Face. Type a Swahili word below and you should see Runyambo come back. For the other direction, open the lab.

Three SKUs

Quality and Lite both do Swahili → Runyambo (nileagi/nileagi-nym-mt · nileagi/nileagi-nym-mt-lite). Reverse does Runyambo → Swahili (nileagi/nileagi-nym-mt-rev). All three live in the nileagi-nym collection. Quality is the default promote for Swahili → Runyambo; reverse is the checkpoint we ship for the other way. There is no English → Runyambo SKU—pivot English through Swahili first.

What the numbers mean

Supervision is dictionary-style: one Swahili surface (often a single lemma or short gloss) maps to one Runyambo surface. Splits are lemma-safe so related glosses do not leak across train and test. Training pairs are not redistributed with the weights.

On the held-out test set, quality and lite both reach chrF2 16.6 for Swahili → Runyambo (identity baseline about 8.0). Reverse reaches chrF2 13.6 (identity about 6.0). A score in the mid-teens on short phrases usually means stems are partially recovered—not reliable dictionary lookups. Word-level BLEU is often near zero even when the stem is recognisable. Treat the release as a research preview.

What to expect

Short lemmas sometimes land near the reference stem while still missing diacritics or final syllables. Extra spaces inside Runyambo words can show up in decode. Full clauses, chat, legal drafting, medical notes, and news are out of scope. Prefer character-level metrics when you score new text, and keep a human lexicographer in the loop before you treat an output as authoritative.

Applications

This preview is not a production translator for paragraphs. It is a way to put Runyambo lemmas and short phrases beside Swahili on machines people already have.

  • Lexicon and learning

    Look up a Swahili word in Runyambo—or the reverse—then edit with a speaker before it becomes teaching material.

  • Community wording

    Draft short glosses for notices, radio scripts, or word lists when Swahili is the bridge language and Runyambo is the local surface.

  • Low-resource research

    Reuse the open Runyambo checkpoints as a starting point for Bantu lexical MT where sentence bitext is still scarce.

For builders

For Swahili → Runyambo, set source swh_Latn and force BOS nym_Latn. For Runyambo → Swahili, set source nym_Latn and force BOS swh_Latn. Omitting either side of the pair will not match the reported scores. The lab and the blog try box already do the right tags for you.

Open weights

Quality, lite, and reverse on Hugging Face. Methods and limits are in the technical report that ships with the collection.

Collaboration

If you are extending this work, evaluating on new glosses, or building with Runyambo in your community, write to us at hi@nileagi.com. The weights are released under CC BY-NC-SA 4.0; commercial use needs a written agreement with NileAGI. Training data is not redistributed with the checkpoints.

Full preview

The Runyambo Machine Translation preview in the lab adds both directions in one form—beyond the blog box above.

Go to preview