Resep ID Gemma 4 E2B IT - Indonesian Recipes Model
Google's Gemma 4 is fluent in Indonesian, but when asked to write a recipe for a specific dish like Tengkleng Kambing or Tahu Tek Khas Surabaya, it often produces a generic Indonesian recipe instead. The base model scored just 2.7/5 on dish authenticity. It knew the language but not the food. We fixed this by fine-tuning Gemma 4 on 66,000 real Indonesian home-cooking recipes using an AMD MI300X GPU. One epoch of full supervised fine-tuning was enough to teach the model which ingredients and steps belong to which dish. The result: dish authenticity improved from 2.7 to 3.25, format compliance jumped from 3.9 to nearly 5.0, and the model now consistently produces properly structured recipes with correct Bahan/Langkah formatting. The fine-tuned model was quantized to GGUF so it runs on any laptop, available on Hugging Face.
Category tags:
For Innovators & Creators