[anon373839]

From the limited testing I've done, Gemma 3 27B appears to be an incredibly strong model. But I'm not seeing the same performance in Ollama as I'm seeing on aistudio.google.com. So, I'd recommend trying it from the source before you draw any conclusions.

One of the downsides of open models is that there are a gazillion little parameters at inference time (sampling strategy, prompt template, etc.) that can easily impair a model's performance. It takes some time for the community to iron out the wrinkles.

[moffkalast]

At the end of the day it doesn't matter how good it its, it has no system prompt which means no steerability, a sliding window for incredibly slow inference compared to similar sized models because it's too niche and most inference systems have high overhead implementations of it, and Google's psychotic instruct tuning that made Gemma 2 an inconsistent and unreliable glass cannon.

I mean hell, even Mistral added system prompts in their last release, Google are the only ones that don't seem to bother with it by now.

[hnfong]

If you actually looked at gemma-3 you’ll see that it does support system prompts.

I’ve never seen a case where putting the system prompt in the user prompt would lead to significantly different outcomes though. Would like to see some examples.

(edit: my bad. i stand corrected. it seems the code just prepends the system prompts to the first user prompt.)

[CoryLR]

This has been my experience as well. I don’t use system prompts anymore. Curious if there’s a good reason to start.

[moffkalast]

It matters as a common standard for model integration. System messages aren't removed when regular context messages start to get cut, the model pays more attention to tools and other directives defined there. Like situational context, language it should use, type of responses, etc. OAI/Anthropic give their models system prompts a mile long to tune their behaviour to a T.

There's also the ideological camp with Hartford and Nous and the rest where models are supposed to be trained as generally as possible, with the system prompt being strictly followed to adapt it to specific use cases.

[tobacco-sausage]

Direct prompt injection attacks can be somewhat mitigated by using system prompts (though not completely) if that helps.

[moffkalast]

I've read the Gemma 3 technical report, it doesn't mention anything about it in the format section. Did they forget to include that? Where did you find the source that claims otherwise?

[staticman2]

Google AI Studio offers no system prompt option for Gemma 3 which also suggests this model doesn't support it...

[moffkalast]

The official ggufs have this in them as an example, I'm even more confused now:

<start_of_turn>user

You are a helpful assistant

Hello<end_of_turn>

<start_of_turn>model

Hi there<end_of_turn>

<start_of_turn>user

How are you?<end_of_turn>

<start_of_turn>model

[hnfong]

I was confused by this too, but now I think they’re just suggesting to prepend the system prompt to the first user prompt.

I think the ggufs took the chat template defined in tokenizer_config.json , which basically does that.

[hnfong]

My bad, original comment updated.