Kenny0bi/ami

Yoruba diacritic restoration: a 1.29M-parameter CPU model that puts the marks back. 92.4% char accuracy on MENYO-20k. # ami A Yoruba diacritic restoration model, trained on a 2014 laptop, published on the Hugging Face Hub. **The model: huggingface.co *àmì* is Yoruba for "mark". This is the most personal project in my portfolio. I grew up speaking Yoruba in a village in southwestern Nigeria, and written Yoruba depends on marks that most keyboards drop: the dotted letters ẹ, ọ, ṣ are different letters entirely, and the tone marks separate words that are otherwise spelled the same. Type `oko` and you have collapsed oko (farm), ọkọ (husband) and ọkọ̀ (vehicle) into one string. Most Yoruba text on the internet lives in that collapsed state. This model puts the marks back: 1.29 million parameters, running at 9,600 characters per second on a twelve-year-old CPU, because a tool for restoring text messages should not need a GPU. ``` in: bawo ni oko re se n lo out: báwo ni ọkọ rẹ̀ ṣe ń lọ (how is your husband doing) ``` It chose *husband*, not *farm* or *vehicle*, from context. That choice is the entire project. ## Results Evaluated on the test split of MENYO-20k (Adelani et al. 2021), the professionally translated and diacritized benchmark: 6,573 sentences never seen during training or model selection. | system | char accuracy | word accuracy | ambiguous words | sentence exact | |---|---|---|---|---| | leave text unmarked | 39.2% | 15.5% | 12.2% | 0.1% | | word lookup (most frequent form) | 77.1% | 69.2% | 78.6% | 3.8% | | **ami** | **92.4%** | **86.0%** | **88.2%** | **10.6%** | Char accuracy is scored only on positions that can legally carry marks, so nobody gets credit for copying consonants. The margin over the lookup table lives where it should: on the ambiguous words, where the same spelling has several attested forms and only context can decide. ## The information story Before training anything, I measured what the marks actually carry. Knowing only the base letter, a mark costs **1.64 bits per character** to guess; *e* and …