Open research
Everything MEYNG builds for Sango is published, with its evaluation attached — including the results that do not flatter us.
Benchmarks
Out-of-distribution, against the untrained NLLB-200 baseline, beam 4. Neither test set was used to select the model. chrF is primary; bare BLEU is unreliable for Sango morphology and should not be quoted alone.
| Direction | Test set | This model | Baseline | Verdict |
|---|---|---|---|---|
| Sango → French | FLORES+ devtest | 31.88 | 31.37 | We win (+0.51) |
| Sango → French | SIB-200 test | 31.45 | 30.20 | We win (+1.25) |
| French → Sango | FLORES+ devtest | 32.56 | 33.70 | Baseline wins (−1.14) |
| French → Sango | SIB-200 test | 34.32 | 36.03 | Baseline wins (−1.71) |
Verified against Hugging Face, 6 Sept 2026.
chrF on FLORES+ devtest and SIB-200 test, beam 4, transcribed from the model card's own benchmark tables. Nothing here is estimated.
Retrieved 6 Sept 2026
We recommend the untrained baseline for French to Sango and publish the number that says so. Route by direction.
Open research
241
Downloads of the open Sango vocabulary dataset ({recent} in the last 30 days)
Verified against Hugging Face, 10 Oct 2026.
The Hugging Face datasets API 'downloadsAllTime' field, cumulative since the dataset was published in April 2026. The figure in brackets is the 'downloads' field, which Hugging Face defines as the previous 30 days.
Retrieved 10 Oct 2026
231
Downloads of the Sango to French translation model, last 30 days
Verified against Hugging Face, 10 Oct 2026.
The Hugging Face models API 'downloads' field, which Hugging Face defines as the previous 30 days. Not cumulative.
Retrieved 10 Oct 2026
Sango to French translation model
Fine-tuned from NLLB-200. Beats the baseline out of distribution in the Sango to French direction on two independent benchmarks; loses to it in the other direction, which the model card states plainly.
Research use only, CC BY-NC 4.0, inherited from the base model.
Sango vocabulary dataset
Trilingual Sango, French and English entries with per-entry provenance and review status. Entries that cannot be verified are withdrawn, and the audit trail is public.
CC BY-SA 4.0. Commercial use allowed.
sango-nlp toolkit
Tokeniser, stemmer and language detection for Sango, on npm.
MIT.
Hugging Face
Every model and dataset MEYNG has published, in one place, with their evaluations and licences.
Licences, plainly
The line between free and paid is drawn by licence, not by strategy. Here is exactly where it falls.
The translation model is CC BY-NC 4.0, inherited from NLLB-200 and not removable. Research and evaluation only — it cannot be licensed for commercial use in this form, by us or anyone else. If you need Sango translation in a commercial product, that runs through the API, which is built on different foundations.
The vocabulary dataset is CC BY-SA 4.0 and the toolkit is MIT. Both allow commercial use. Share-alike applies to derivatives of the dataset.
The corpus
The public web-mined French–Sango corpus advertises 588,039 quality-filtered pairs. A forensic cleaning pass found 460,437 duplicate Sango targets, 100,683 rows carrying leading verse numbers and 29,939 carrying speaker labels. The honest unique corpus after cleaning is 45,451 pairs. Earlier models trained on the raw corpus learned those artefacts and regressed below the baseline out of distribution; that is why every release now passes an out-of-distribution gate before it ships.