Test Gemma 270M and Gemma 3n

[supreme-gg-gg]

This issue is about testing our current platform on these two new Gemma models because most development validation happen on 1B and 4B text and vision. The final goal is to document these on the project docs as well! 😄

270M

This model is built with the vision of fine tuning and task specific application -- the right tool for the right job. Might be our most-recommended model for a lot of users!!

Test the usage of this model on all our methods (including dataset modality, types, trainer, frameworks, backends, etc.) to make sure this always works and is blazing fast with fine tuning!

No changes should be required except adding this to the list of valid models to choose from.

Good candidate use case: more instruct based text dataset, SFT should definitely work, not sure about preference and reasoning though need some research to back it up

3n

3n has an unique audio modality support, but unfortunately we do not support it right now, we hope to bring it in very soon!! However, for now we should still validate the usage for text + vision. Again, should work OOTB because it's just 1B and 4B training process.

Would be cool to test GRPO and DPO in addition to SFT for this one.

[inclinedadarsh]

On it!

[inclinedadarsh]

So I tried creating a new job. Could you elaborate on this part, please?

including dataset modality, types, trainer, frameworks, backends

I understand that we can test it on either text or image modality. After that I am not sure what you mean.

In the frontend I saw there are multiple training methods, so that's one thing. Other than that, do we have different trainers and frameworks and backends as well in the configuration now? Or do you mean to test it on Notebook first, if it works fine then move it to the application?

[supreme-gg-gg]

@inclinedadarsh for what modality types etc mean checkout the jobs/training-job/README.md in the RL branch they refer to how you can configure a job by setting like trainer, PEFT, unsloth vs hf, export options, etc. However, since the RL training image is not in prod yet (i.e. this is implemented but not shipped yet).

For now let's keep just test it on a multimodal dataset using unsloth (i.e. you can ignore these config and just use the frontend), since trainer is not supported yet. I will update on this, sorry for the delay

[supreme-gg-gg]

For this issue let's try this dataset: Thytu/ChessInstruct from the latest unsloth 270M notebook: https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Gemma3_(270M).ipynb

Seems to work OOTB 😄

[supreme-gg-gg]

For 3n:

1. Text: https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Gemma3N_(4B)-Conversational.ipynb, they're using this dataset: mlabonne/FineTome-100k, however it's preformatted in ShareGPT 😭 😭 we cannot use it because we don't support conversion from ShareGPT to ChatML yet, we only support formatting ChatML (not conversion)

2. Vision: https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Gemma3N_(4B)-Vision.ipynb, they're using the latex OCR dataset -- we already have that for 4B

[inclinedadarsh]

For the first one, we can format the dataset locally and then test the model using it?

Also, in the notebook I saw this:

from unsloth.chat_templates import standardize_data_formats
dataset = standardize_data_formats(dataset)

We can use this? Or maybe just write our function and use dataset.map lol.

[supreme-gg-gg]

For the first one, we can format the dataset locally and then test the model using it?

Also, in the notebook I saw this:

from unsloth.chat_templates import standardize_data_formats
dataset = standardize_data_formats(dataset)

We can use this? Or maybe just write our function and use dataset.map lol.

Ah yes this is already there in the training code, we standardise it before training. However the processing service doesn't support ShareGPT 😭 so that sounds like a design limitation, we should put up a feature request. For now we can pick a new dataset or just format locally