The project had no translation model yet. NLLB-Maloba-h1 is the full fine-tuning of NLLB-200 distilled 1.3B, Meta's multilingual model, on 288 French → Lingala pairs translated by Lingala speakers and harmonised.
The result
On the internal test set (79 sentences), BLEU rises from 23.37 to 32.00. On FLORES+, the external reference of 1,012 sentences, it rises from 15.49 to 16.39. Both gains are significant (p ≤ 0.001, paired bootstrap of 1,000 draws, corrected for multiple tests).
The lesson of the 288 pairs
A second version, h2, added 4,700 pairs built by automatic back-translation. It brings nothing on the internal test and falls back on FLORES+. It was rejected.
It is human translations that move Lingala forward. Every sentence you translate counts.
The model inherits the CC-BY-NC 4.0 licence of NLLB-200: it may be used for research and non-commercial purposes. Three validators judged 38 translations: 66 % correct, 97 % correct or recoverable. Their agreement is weak, so this result remains indicative.
By The Maloba team, September 26, 2026.