3 posts on Fine-tuning.
The raw 27-cell Kaggle notebook for the Naamapadam PII fine-tune. What eleven Indian languages of NER data actually gave me.
A Gemma 4 fine-tune that rewrites discharge summaries at Flesch-Kincaid grade 3.8 while holding 94% information preservation. Deployed as a Hugging Face Space.
The multilingual model was the obvious pick and it lost by six F1 points on Bundesdatenschutzgesetz entities. Tokenizer coverage beat parameter breadth on German.
