JosueG/adja-nmt-nllb-600m-forward-r10ks4k-seed42

NLLB-600M French -> Adja

This is the thesis-final French-to-Adja machine translation checkpoint for the May 2026 CS thesis release. It is one of the two NLLB models used to make the French-Adja MT work inspectable and reproducible.

Thesis Role

This model is the forward MT artifact: French source text to Adja target text. It supports the thesis experiments on grammar-guided dataset construction, low-resource NMT fine-tuning, and the speech-to-text-to-speech pipeline.

Model And Data

• Task: machine translation, French -> Adja
• Base model: facebook/nllb-200-distilled-600M
• Direction: fra_Latn -> aj_Latn
• Training data: French-Adja thesis corpus lineage, with the public canonical MT reference at JosueG/french-adja-parallel-corpus
• Release repo: FrejusGdm/cs-thesis-may-2026

Headline Result

The seed-42 NLLB-600M full/module runs are the strongest MT baseline in the release. The public GitHub result tables report about 21-22 BLEU and 29 chrF on the held-out MT evaluation for the full-corpus NLLB setting.

See:

• results/mt/summary/all_results.csv
• results/mt/README_thesis_qualitative_decode.md

Limitations

• Adja remains low-resource; outputs should be reviewed before community, educational, or publication use.
• Metrics are sensitive to corpus split, normalization, and reference coverage.
• The model uses the release tokenizer/config; load this repo directly so the Adja language token and generation settings match the checkpoint.

Citation

If you use this model, cite:

Josue Godeme. 2026. CS Thesis May 2026: French-Adja MT and Adja Speech Experiments. https://github.com/FrejusGdm/cs-thesis-may-2026