MICWEN PRO

Exactly right on the bootstrapping loop — that's precisely the progression we're running.

Small precision on the mechanism: the model has seen some Sango during pretraining (it appears in Common Crawl), but not enough to produce coherent translations cold. The vocabulary injection doesn't teach the language from scratch — it gives the model enough anchoring signal to activate what it weakly learned. The grammar rules and orthography notes handle the parts pretraining didn't cover reliably (tonal distinctions, diacritics, Sango-specific syntax).

And yes, the loop you're describing is live: the vocabulary-augmented outputs → native-speaker verification → parallel corpus → fine-tuned NMT model. We just published BENCH-001 results on the fine-tune: +5.70 BLEU over baseline on French→Sango, +9.10 on Sango→French. The vocabulary-augmented prompting approach (BLEU 2.92 on the same task, zero fine-tuning) is the floor; the fine-tune is what you get once the dataset is big enough.

The data pipeline post documenting that second step just went up here: https://huggingface.co/blog/MEYNG/sango-vocabulary-pipeline

The interesting open question is where the ceiling is for a 600M-parameter model on a language with ~5M speakers and sparse digitized text. We're nowhere near it yet.