google

/

shieldgemma-2-4b-it

like 134

Follow

Google 35.5k

@BITDDD you can probably fine-tune it on different images and policies like how you would fine-tune a vision language model.

Use Unsloth for gemma3 and vor visual fine tuning is working great.

Hi , Apologies for the delay,

Yes, you can extend the model's capabilities to recognize new concepts and tags, but it's done through a process called fine-tuning, not through a fully autonomous, self-training loop.
Gemma 3 (if you're referring to the multimodal variant) can be fine-tuned like other vision-language models (VLMs), but tag extension and image-specific training depend on how your training data is structured and what framework you're using.

Kindly refer this documentation for more information. Thank you