Finetuning Guide
Is it possible to get the fine tuning sample notebook to better understand the finetuning process for this model. if yes kindly share the notebook
The pre-converted models we have so far are "-web.task" format, which we don't have any fine-tuning notebooks or colabs for, and probably won't be able to make any time soon. But you have two options for further fine-tuning:
You can try to use a "non-web" fine-tuning colab/notebook to get a "non-web" .task file. If the model is small enough (1B or smaller) than this can potentially run on web too, but this will almost certainly fail for 4B or larger models, like TranslateGemma. We have colabs like this for Gemma 3 270m, for example.
Alternatively, you can use external tools to fine-tune the model directly from the source google/translategemma-4b-it, keeping it as a .safetensors model. Then you can convert to our "-web.task" format using the "MediaPipe LLM Converter". Note that most of the documentation on our website for model conversion will point you to a different converter which will not work for this purpose.
The pre-converted models we have so far are "-web.task" format, which we don't have any fine-tuning notebooks or colabs for, and probably won't be able to make any time soon. But you have two options for further fine-tuning:
You can try to use a "non-web" fine-tuning colab/notebook to get a "non-web" .task file. If the model is small enough (1B or smaller) than this can potentially run on web too, but this will almost certainly fail for 4B or larger models, like TranslateGemma. We have colabs like this for Gemma 3 270m, for example.
Alternatively, you can use external tools to fine-tune the model directly from the source google/translategemma-4b-it, keeping it as a .safetensors model. Then you can convert to our "-web.task" format using the "MediaPipe LLM Converter". Note that most of the documentation on our website for model conversion will point you to a different converter which will not work for this purpose.
Yeah Thank you