Hi There! Could you tell me how to convert 'Gemma3' to text only model?

#1
by CocoRoF - opened

Hi!
First of all, thanks for sharing this attempt.
If you don't mind me asking, could you tell me how you made it a Text Only model?
Also, have you ever tried to see if you can effectively tune only the Text part in the same way as Lora, and if so, could you tell us about your experience?
Thanks!

Owner

Hi! No problem.

Could you tell me how you made it a Text Only model

Here's the script: gapeleon/gemma3-novision-converter

4b and 12b work on the A100 40GB in colab. 27b needs more VRAM.

have you ever tried to see if you can effectively tune only the Text part in the same way as Lora

Yeah, that's actually how I made this test model:

gghfez/R1-Gemma-3-4B-multimodal-test

  1. Removed the vision adapters
  2. LoRA training on gemma3-4b-novision
  3. Put the vision adapters back + tested

Of course since then, Unsloth have added support for the vision model, so there's no need to do this just for SFT.

Thanks for getting back to me!
Your kind answer was super helpful, and I really appreciate you sharing your knowledge.

Hi bro, I find the text-only versions very useful, but the link to your script seems to be not working.
May I kindly ask you to upload the script again, or also provide the converted models for "pt" (pretrained) gemma-3 models as well?
Many thanks!

Sign up or log in to comment