Fix the API. Training, streamGenerateContent

Gemini API
1

  1. Further training is broken. Used baseModel instead of tunedModelSource.tunedModel.

  2. CreateTunedModelRequest.tuned_model: text_input is too long. The maximum character count accepted is 40000.
    CreateTunedModelRequest.tuned_model.tuning_task: output is too long. The maximum character count accepted is 5000.
    It counts bytes, not characters. JSON is UTF8, not ASCII. And that’s not enough. For a 16k model(flash), 16k*4=64k characters are required.

  3. Add tunedModels.streamGenerateContent.

  4. Incorrect counting of epochs in operation metadata.snapshots.

2

It’s actually counting tokens, which are pretty complex, and there is no easy way to make sure that all 64,000 characters fit within the context, which is probably why they have a max character count of 40,000 to stop all major errors regarding this.

You can check out more info on it here: Understand and count tokens  |  Gemini API  |  Google AI for Developers

I am not sure if there is a feature at the moment that allows you to be able to fine-tune a previously fine-tuned model. This might not be a bug, but I can’t seem to find anything in the documentation regarding this.

Do you have any examples or more information you can share regarding this?

3

Nothing has changed in the past two months.
The developers don’t seem to be looking at the forum.