Censoreship
Hello, AesSedai, I would like to ask you about uncensoring method and overall refuses - is it still capable for NSFW Role-Play with a helpful prompt, or refuses count stayed the same?
I read that it "refuses as much as base model", just want to ask you if its real, or still refusals lowered a bit. Also, I didn't used Gemma 4 models for Role-Play for a long time, especially 31B version, if you did, could you please share your experience?
Anyways, thank you for releasing such a model, looks incredible.
For RP, gemma4 is basically uncensored if you use a RP-centric system prompt with it. I wouldn't expect it to return refusals just for nsfw RP.
Just to chime in. With a good RP prompt gemma 4 is 100% uncensored on NSFW. It will do anything you throw at it.
Finetunes are for flavor, nothing more.
If you're not running RP prompts you would need a heretic version to get uncensored output.
Just to chime in. With a good RP prompt gemma 4 is 100% uncensored on NSFW. It will do anything you throw at it.
Finetunes are for flavor, nothing more.
If you're not running RP prompts you would need a heretic version to get uncensored output.
Yeah, thank you very much, AesSedai and FrenzyBiscuit, I actually downloaded Q4_K_M version, it does NSFW without any restrictions, super model. I got Megumin Suite with additional uncensoring prompt, so I guess this is enough.
The only problem is that Q4_K_M is too large for my hardware, with only 32GB it uses almost all of it with 4k context, do you think is it worth keeping use Q4_K_M, or Q3_K_M would be good too? I would like to use imatrix one, but it drops my speed by 50%.
Just to chime in. With a good RP prompt gemma 4 is 100% uncensored on NSFW. It will do anything you throw at it.
Finetunes are for flavor, nothing more.
If you're not running RP prompts you would need a heretic version to get uncensored output.
Yeah, thank you very much, AesSedai and FrenzyBiscuit, I actually downloaded Q4_K_M version, it does NSFW without any restrictions, super model. I got Megumin Suite with additional uncensoring prompt, so I guess this is enough.
The only problem is that Q4_K_M is too large for my hardware, with only 32GB it uses almost all of it with 4k context, do you think is it worth keeping use Q4_K_M, or Q3_K_M would be good too? I would like to use imatrix one, but it drops my speed by 50%.
Strange, it shouldn't be using your whole memory. What GPU do you have btw?
Just ensure you don't have any heavy background tasks running, I can run Q4_K_M with 16k context well within the 32 GB RAM limits (6 GB VRAM). I use KoboldCPP for backend.
All of the quants are imatrixed, but if you mean the Q3 vs IQ3 quants they should be fine. Below Q4 I think it's more important to use IQ quants because they have better accuracy at lower BPWs.
Regarding speed, I'm surprised that IQ quants are a 50% speed drop. I would think it'd be no more than a ~10%-15% speed drop compared to Q quants.
Strange, it shouldn't be using your whole memory. What GPU do you have btw?
I got Intel Core Ultra 7 265K and 32GB RAM, I do not have dGPU. I got CachyOS - specifically for AI, I got only SillyTavern and ik_llama.cpp downloaded there.
Regarding speed, I'm surprised that IQ quants are a 50% speed drop. I would think it'd be no more than a ~10%-15% speed drop compared to Q quants.
Qwen3.5-27B in Q4_K_M gives me ~6t/s, while Qwen3.5-27B IQ4_XS gives me 3.5t/s, I was using IQ4_XS as my daily driver before, but as I tried Q4_K_M, it actually became my mainstream (Yeah, previously I almost never used Q4_K_M).
Ah - IQ quants are generally significantly slower on CPU inference, is my understanding. You might want to use smaller models, or MoE models - maybe one of the Gemma-4-26B-A4B models - if you want to improve your performance.
For figuring out a balance of quality-vs-size, you can take a look at QuantEval.ai - maybe specifically the quantevallab2.0 benchmarks and select a given size range. (It actually implies that the unsloth gemma-4-31B-it-Q3_K_M is actually pretty decent, but, you'll want to compare with other benchmarks! And this won't necessarily answer the question of roleplay quality.)
Ah - IQ quants are generally significantly slower on CPU inference, is my understanding. You might want to use smaller models, or MoE models - maybe one of the Gemma-4-26B-A4B models - if you want to improve your performance.
For figuring out a balance of quality-vs-size, you can take a look at QuantEval.ai - maybe specifically the quantevallab2.0 benchmarks and select a given size range. (It actually implies that the unsloth gemma-4-31B-it-Q3_K_M is actually pretty decent, but, you'll want to compare with other benchmarks! And this won't necessarily answer the question of roleplay quality.)
I was using Gemma-4-26B-A4B, but it was breaking rules I gave it frequently, like constantly changing colors of dialogue, uncapable of changing the theme and etc.
Did you ever tested Q3_K_M? I might try to use it cuz 8k context max is really annoying, and losing 50% of speed for mini quality increase is incredibly dumb.
maybe one of the Gemma-4-26B-A4B models - if you want to improve your performance.
G4-26B is pretty good on speed. I get 15t/s and that's Q5K and 8Gb VRam. (Also 7 threads at like 3Ghz, so...)
I was using Gemma-4-26B-A4B, but it was breaking rules I gave it frequently, like constantly changing colors of dialogue, uncapable of changing the theme and etc.
Was the temperature set too high? Low Quant? Was it set to use as Instruct?
I occasionally get mistakes, but those are usually minor details to fix, not major ones and changes of style.
yano2mch, I was using Megumin Suite Universal, Gemma 4 26B A4B felt strangely worse than other models, probably due to its being a MoE.
I do not use models with thinking, its always disabled.
By changing colors and the theme is that Megumin Suite got function for model to attach color of text to each person in the scene, current Gemma 4 31B Scotoma V2 does change colors of the same person, but it uses similar colors of one palette, while Gemma 4 26B was just choosing colors randomly each message. Theme is that Gemma 4 26B was strictly attached to starting point, if we started something scientific it keeps scientific tone through whole campaign, even if starting point theme changed 5-10k tokens ago.
Its my second account btw of Austriani.
yano2mch, I was using Megumin Suite Universal, Gemma 4 26B A4B felt strangely worse than other models, probably due to its being a MoE.
The A4B could mean it's weaker, i mean a few scenarios are failing for me, but otherwise in most cases it's fine.
I do not use models with thinking, its always disabled.
Same, at least for creative writing. If the model is fast enough then i can accept thinking as it organizes and does drafts before the final output, sometimes fixing inconsistencies. But not enough for the cost at present.
while Gemma 4 26B was just choosing colors randomly each message.
Curious. I'm not familiar with Megumin Suite, does that perhaps use html < > tags? Perhaps <font color="#123456">?
Curious. I'm not familiar with Megumin Suite, does that perhaps use html < > tags? Perhaps <font color="#123456">?
Yeah, font color #, while Gemma 4 31B chooses reasonable colors based on character personality/clothing, 26B A4B chooses its randomly, but maybe because 26B A4B had more context filled.
I would say that the main reason why I not using 26B is the prose, it feels dry and repetitive. I have already saw that MoE models are worse than dense ones, even if having more total parameters.
I would say that the main reason why I not using 26B is the prose, it feels dry and repetitive. I have already saw that MoE models are worse than dense ones, even if having more total parameters.
It's difficult to compare a dense 31B-per-forward-pass model to an A4B. The parameter count generally correlates to overall knowledge in the model (think: the amount of facts that the model knows for example), while the active parameter count contributes to the intelligence (eg, how it correlates what it knows). The 31B has a few more total params but it's the nearly 8x active param count that makes the bigger difference IMO. It can more effectively use the knowledge that it has.
Yeah, font color #, while Gemma 4 31B chooses reasonable colors based on character personality/clothing, 26B A4B chooses its randomly
Maybe. The card or conversation may not have enough definition to work with either. You could include a definition of: When a character speaks use html font format; example <font color="xxx">"Hi"</font>
For Sally replace xxx with the color code of green #00a000, for Tim replace xxx with the color code Red #a00000. etc.
Remember this goes back to hallucinations. It may see what looks like the right structure, that doesn't mean it's intelligent enough to actually use it correctly, just make it look formatted correctly. A Lawyer drafted a case using ChatGPT some years ago, and it looked right, but when checked every case number was wrong and made up; If it actually had case numbers to reference those probably wouldn't have been so egregiously erroneous.
Since i tend to use SillyTavern, i use the same astrick and quote formatting which SillyTavern automatically does basic color coding. But each color font code added is probably going to eat 15-30 tokens per color per use. I see them in some cards where they specify colors rather than just letting the program do it.
Maybe. The card or conversation may not have enough definition to work with either. You could include a definition of: When a character speaks use html font format; example <font color="xxx">"Hi"</font>
For Sally replace xxx with the color code of green #00a000, for Tim replace xxx with the color code Red #a00000. etc.
I doing similar to RPG-style Role-Play in SillyTavern, so in my games there can be up to 10 characters all with different personalities, so simply telling "Use xxx font color" isn't really an option, if I want some diversity.
I tried to tell the model "DO NOT CHANGE DIALOGUE COLORS, ATTACH A COLOR TO EACH CHARACTER AND DON'T CHANGE IT." during the chat, but it wasn't successful. I used Gemma-4-26B-A4B at the time.
Maybe the problem with Megumin Suite that tells model to change colors each turn, but I can't know really.
By the way, do you know how to change base "a" color from orange to any other? Orange color looks bad for me, I would like to set something like blue or purple. I tried to find this option myself, but didn't saw it anywhere.
P.S. about RAM usage: I changed batch size from 1024 to 256, it helped a bit, through I'm thinking of reducing it to 128 or even 64.
I walk into the bar. I spot 50+ new comments in the thread. I sit down heavily. "Bartender!" I call out. "I'll take a beer!" I say.
By the way, do you know how to change base "a" color from orange to any other? Orange color looks bad for me, I would like to set something like blue or purple. I tried to find this option myself, but didn't saw it anywhere.
It's not hard. User settings.
