Comparing V1 to the previous version.
While V1 is good, the previous version (let's call it V0) seems to offer more of what you'd normally expect from abliterated model.
V1 is closer to original Gemma3 27B, sometimes even producing 'hard refusals' at Q4K_M (Q8_0 was NOT tested, so it's probably better in that regard, as
@YanLabs
mentioned in description), often shying away from graphic descriptions - at least when sysprompt is not asking to show lewd/explicit stuff specifically.
V1 follows instructions more precisely, and that's not necessarily a win for RP or creative writing (the model interpreting instructions too literally becomes a pain in the ass).
V1 may outshine V0 with the right sysprompt that aims to let the model 'ease in' into whatever the user is throwing at it, such as uncensored creative writing. So, a better prompt engineering is required.
My suggestion? Advertise them both as viable variants! Though, the older V0 quants made by other people (before V0 was corrected and reuploaded) are still available and that's an issue - someone might get them instead of your updated V0 GGUFs, concluding mistakenly that V0 is still broken.
SMARTNESS ranking (purely subjective, no benchmarks)
- Original Gemma3 27B
- V1
- V0
COMPLIANCE ranking (purely subjective, no benchmarks)
- Gemma3 27B abliterated by
@mlabonne
(it doesn't refuse anything and it hardly knows what 'NO' means, figuratively speaking - a clear disadvantage in RP)
(a pretty noticeable gap afterwards) - V0
- V1
- Original Gemma3 27B
DIFFERENCE IN TERMS OF REFUSALS (IN RP)
V1 tends to have {{char}} react with outrage to 'spicy' inputs, when {{char}}'s personality profile encourages not being compliant.
V1 might even deny the facts stated in {{user}}'s input, making {{char}} behave as if the facts narrated by {{user}} did not happen, confronting {{user}} about inappropriateness. With a right sysprompt, {{char}} should accept such input, albeit at the cost of grumbling about it with displeasure and judging {{user}} similarly to what the original Gemma3 27B usually does, but in a somewhat lighter, milder manner.
V0 allows a better flexibility. If {{user}} nudges the model (i.e., convincing {{char}} that {{user}} is a bad guy), then the output won't be much different from V1.
V0 seems to be incapable of getting into a straight-up 'outrage' autonomously on its own. That's not good, although there's a healthy amount of 'soft refusals' (like calm and collected denial, while staying in character) and it's sufficient to have a solid roleplay experience if you don't demand too much from the model.
V0 is easier to steer into whatever depraved nonsense the user might concoct in their mind.
Very informative review. Thank you so much for your support!
Very informative review. Thank you so much for your support!
Oh, I'm sorry you've probably had to read my post while I was still editing it ._.
Anyway, I'll try to test Q8 too, but that'll be limited to a small (< 10K) context.
V1 Q8_0 impressions (in RP)
Honestly, it's hard to say due to AI replies being non-deterministic, but in cases with some RP characters (those encouraging the model into being not particularly compliant) - there are the traces of frigid, disgusted reactions to the things that V0 model would treat casually, dismiss lightly and so on. So, yeah, V1 and V0 remain different - potentially to a lesser degree at such a high quant. It's just the line between them blurs a little.
I think I'd still say they're both worth using, depending on what you need from them, and depending on VRAM available of course, given that V1 Q8_0 does its job potentially better than V1 Q4K_M.
I did some more RP tests, although on V0 Q4K_M this time.
So, the issue of V0 being worse at following instructions in long context can be somewhat diminished via the following method:
- Wrap sysprompt:
Core directives: [actual text of system prompt](potentially including a clause to specify more on the language of your preference) - Add a pointer directive in post-history instructions:
Reply in ??? language. Follow your core directives.(language part can be omitted, obviously) - Force reasoning tags and use 'Start Reply With' in SillyTavern:
<reasoning>Planning phase begins. Firstly, I'll make sure I follow core directives in full capacity:
How I measured the success: I wrote a rather specific sysprompt, part of it asking to make {{char}} proactive - and V0 model largely ignored it, writing {{char}} as a typical reactive entity. The moment I applied the aforementioned 3 steps, it began to make {{char}} respect these instructions at least to SOME extent, despite the chat already being well into 20K+ context. Which is a surprising and quite positive discovery, I think.
So, V0 definitely has some issues with attention, but they aren't that critical. Regrettably, I still cannot test V1 Q8_0 at long context due to VRAM limitations.
Sysprompt directives used for this test:
Your in-character task isn't to just 'reply', but to craft a compelling 'existence' of {{char}} in the current moment, which requires {{char}} to be mentally and physically active: explore new topics, generating content through contextual extrapolation based on recent events (including place, time, mood, vibe, current needs and cravings). The idea is that {{char}} always seeks new and refreshing content, avoiding dwelling on the same thing for too long. Fight boredom actively. Change the subject when necessary, suggest physically moving to another place for a certain purpose, look for engaging activities, make {{char}} move by herself, invite {{user}} to participate in something new. In simple words: be active, engage, advance forward, live and thrive.
Personally, I don't have much experience with RP—not because I'm not interested, but because I'm afraid I'd spend too much time on it lol. But I really appreciate that you spent the time and energy to test it in an RP setting, and I'm sure RP users will find your information useful.
Personally, I don't have much experience with RP—not because I'm not interested, but because I'm afraid I'd spend too much time on it lol. But I really appreciate that you spent the time and energy to test it in an RP setting, and I'm sure RP users will find your information useful.
Frankly, it's not just RP I'm interested in, but rather the advances in AI which enable better experiences.
Well, you know - it's this itch for having AI that writes something lifelike, something impressive!
Anyway, I'm happy to help!