Custom Finetune of Gemma27
Hi OddTheGreat,
First of all great work on Mars model. I have tested a few of Gemma27 model and this is so far the best
I have a fairly specific challenge I’d like to run by you. We’re looking to contract someone with hands-on fine-tuning experience, and we’d of course cover your time as well as any training resources needed.
Here’s the situation:
We’re currently using a fine-tuned version of Llama 3.3 70B. It has an excellent writing tone and produces great responses, but it struggles with non-English languages, its attention isn’t great, and on a 5090 we’re limited to running only the 3-bit version.
On the other hand, Gemma 27B addresses all of these issues except one: writing style. Even the fine-tuned versions we’ve tested still missing emotions and sound too much like a generic “helpful assistant.”
We have millions of chat messages between users and Lamma, and our goal would be to fine-tune Gemma 27B on these responses so it can adopt Finetuned Lamma conversation style while keeping Gemma’s strengths.
Is this something you’d be open to discussing in more detail?
Best regards,
Adam
Good time of day.
First of all, thank you for the praise, I am pleased.
Secondly, even if I am open to any discussions, I have to answer you "no". There are several reasons for this, and the simplest one is that I did not fine-tune the models, but only merged them. For now.
I can only advise you to rethink your system prompt used for gemma3, as this greatly affects how emotional the model's responses will be.
Also, I don't want to sound rude, but your message looks suspicious, especially considering your account is empty.
Anyway, thanks again for the positive feedback about the model, have a nice day.
Sure fair enough. Thanks for a quick answer .