k1mb3rlyie/mupun-mt

First Mupun language machine translator # Mupun Machine Translation (English → Mupun) The **first machine translation model for Mupun** (ISO 639-3: `mcu`), a West Chadic (Afroasiatic) language of Plateau State, Nigeria. Mupun is a "Left-Behind" language, absent from NLLB, FLORES, MAFAND, and every major NLP benchmark. This repository provides a working English→Mupun translation model and a simple script to run it. > **Mupun is a distinct language, not a dialect of Hausa or Mwaghavul.** --- ## ⚠️ Honest status: this is an early research baseline Please read this before using the model. **The translations are not good yet.** In native-speaker evaluation by a fluent Mupun speaker, the model averaged about **1.5 out of 5** for adequacy, and performed *worst* on simple everyday sentences (for which almost no written Mupun training data exists) and somewhat better on complex, scripture-like sentences. This model is: - a **proof that Mupun MT can exist at all**, where before there was nothing; - an **honest baseline** to be improved upon; - **not** suitable for real-world translation, and it should not be relied on for any consequential use. The single biggest limitation is data: there is almost no everyday written Mupun to train on. Improving the model is primarily a data problem, not a modelling one. See the accompanying paper for the full evaluation and analysis. --- ## What this model is - **Base:** `facebook/nllb-200-distilled-600M` - **Adaptation:** Low-Rank Adaptation (LoRA), fine-tuned on a multi-domain Mupun–English corpus (New Testament + dictionary + grammar examples). - **Direction:** English → Mupun. - **Language tag:** because Mupun has no tag in NLLB, this model borrows Hausa's tag (`hau_Latn`) as its Mupun target ("Mode A" in the paper). This is why the script forces the target token `hau_Latn` — **that is how this model produces Mupun.** (A dedicated-tag variant, "Mode B," is discussed in the paper; this release uses Mode A, the primary reported system.) A known consequence of the borr …