Fine-tuning method: LoRA instruction tuning via Adaption Labs AutoScientist
Results
Win rate: 69 (fine-tuned) vs. 32 (base model), evaluated on the target dataset
Training converged with stable loss reduction and gradient norm over ~105 steps
Usage
Input: Yorùbá text with diacritics stripped (e.g. Nibi abajade ipade to waye ni ilu Abuja)
Output: Correctly diacritized Yorùbá text (e.g. Níbi àbájádé ìpàdé tó wáyé ní ìlú Àbújá)
License
CC BY-NC 4.0, inherited from the MENYO-20k source corpus.
