So far the best 24B model for RP I've seen until now but, it mixes languages?
I like this model very much since this merge achieves something others cannot, my emotionally complex characters feel more human, it really feels like I'm roleplaying with a human being and not a chatbot however I'm experiencing an issue and I'm not sure of it's because of my samplers, quant, or KV compression. (RP-Spectrum-24B.i1-Q4_K_S running with an early version of Turbo Quant at Q_4)
For instance I doubt the quant or KV compression has anything to do with it, since I run other models like Cydonia 4.1 Base and many others with the same level of quantization and I don't experience this problem.
I roleplay in Spanish, and although some characters use slang, with other models I tried they never blend English words where they shouldn't, for example using Spanglish for markdown actions, it's not a very aggressive contamination, more like... when it happens it throws in 2 to 4 English words inside Spanish replies, something like an English word which failed to fully translate, it's like a degradation lottery, won't always happen but when it does, it breaks my RP immersion. Probably I'm the only one facing this problem since most people will be probably roleplaying in English but it was worth a shot asking.
This small issue is the only problem keeping me from fully enjoying this model at its fullest and I would like to know if there is a way to deal with this because I refuse to just give up on this model, I don't use XTC, adaptive P nor DRY, and TEMP is set at 0.6.
Hi! Sorry for the late response. I had all my focus on another project for a bit.
I roleplay in Spanish, and although some characters use slang, with other models I tried they never blend English words where they shouldn't, for example using Spanglish for markdown actions, it's not a very aggressive contamination, more like... when it happens it throws in 2 to 4 English words inside Spanish replies, something like an English word which failed to fully translate, it's like a degradation lottery, won't always happen but when it does, it breaks my RP immersion.
Since this is a moderate sized merge with a model containing drastic changes to the parameters compared to the other models (Cydonia v4.0) There was always bound to be a good amount of instability.
While instability can have an interesting (in a good way) effect on creative writing models, it can also adversely effect some of the model's knowledge.
Since these are all mainly English-only RP and status tracking trained models, it's no surprise it's Spanish took a hit.
Not really much I can suggest as a fix besides one thing:
Maybe try finding a small yet effective model that you can use for rewriting? Something that's multilingual, but can fit entirely in a CPU?
I believe Google's Gemma-4-E4B and Gemma-4-E2B are multilingual. Using a small model might not seem appealing due to the common misconception that their as dumb as rocks, but they can be pretty handy for instructional tasks (1 optional system role, 1 user role, and an assistant role prefix at the bottom) apposed to conversational tasks (multiple optional system roles, multiple user roles, multiple assistant roles, and any role prefix at the bottom).
If you do decide to go that route, my advice is that you save memory by using a much smaller context window for the smaller model (and enable SWA if you use Gemma4).
It's likely that you can get away with a 2k token limit minimum and not run into any issue.
Greetings, well thanks for taking your time to reply, honestly like I said I like this model enough to not give up on it yet, I've been experimenting with sampler tuning with other models, people sometimes underestimate how much sampler values can affect a model's response, a mistake I was doing quite a lot... I did notice this problem occurs more often with text inside asterisks markdowns only, not with character dialogue lines which is quite interesting, if my sampler tuning approach fail then... here comes option 2...
So yes, I think I know where you're getting at, that idea crossed my mind too, luckily I run this model on a dedicated inference server I have at home so this leaves my main computer's GPU completely free to run a secondary model, I already tried with a very stable 12B model for this, no need for an insane context window when all I need it to do is fix some words in the assistant message, so as you suggest maybe I can lower this value down a lot, 4K or 6K, the secondary model only needs two things... the grammar correction prompt and the assistant message to correct. I did some manual tests in Ooba, it works, the model corrects only the misspelled Spanish words...
My real issue right now is... SillyTavern, I would have liked to use a ST Script with a simple QR Button but... ST is quite stubborn with proper prompt control in the Story String, it also insists on tossing up the entire chat history inside the System Prompt before sending it to Ooba's API and there is no way I know of to deactivate this behavior.
I will probably have to vibecode a simple extension to capture only the AI message and do a grammar correction pass with a secondary model via Ooba's API, I've done it before but for a different approach...
In any case, are there any sampler values you recommend me to use besides temp not higher than 0.7? maybe a json I could try with SillyTavern?
A LITTLE UPDATE: A few hours later after I posted my original reply I began experimenting with samplers (I've been gaining a little more knowledge over time about what they do and how to properly balance an combine them) and I noticed the problem decreased a lot, if not disappeared completely (I don't want to get my hopes up yet, sill doing more testing).
Apparently, having my Rep Pen Range value off the clouds plus an aggressive Rep Penalty.
I did used to have a preset for it, but that is no longer the case. Not to long ago I did a bit of spring cleaning on all my saved presets.
I made this model with the original intention of using it for scene tracking and general backend task, but I've moved on to using MoE models instead for that.
I did used to have a preset for it, but that is no longer the case. Not to long ago I did a bit of spring cleaning on all my saved presets.
I made this model with the original intention of using it for scene tracking and general backend task, but I've moved on to using MoE models instead for that.
I see, interesting... what are the odds, I'm pretty sure I will hold onto this model for quite a long time... as of now the only way I see moving away from it is doing a hardware upgrade and move away to bigger models, yet even some 32B models I tried performed insanely worse than this.
What I'm getting at is, thank you very much for this great model, It will keep me entertained with my role playing sessions with the level of immersion I enjoy, this one did really hit the spot.