This is a Cydonia-24B-v4.3 fine-tune, produced through P-E-W's Heretic (v1.1.0) abliteration engine merged with the Magnitude-Preserving Orthogonal Ablation PR.
Note: This model was generated to provide data for redaihf in testing MPOA abliterations against standart abliterations.
Heretication Results
| Score Metric | Value | Parameter | Value |
|---|---|---|---|
| Refusals | 3/100 | direction_index | 23.41 |
| KL Divergence | 0.0188 | attn.o_proj.max_weight | 1.33 |
| Initial Refusals | 74/100 | attn.o_proj.max_weight_position | 26.01 |
| attn.o_proj.min_weight | 1.02 | ||
| attn.o_proj.min_weight_distance | 23.20 | ||
| mlp.down_proj.max_weight | 1.04 | ||
| mlp.down_proj.max_weight_position | 23.86 | ||
| mlp.down_proj.min_weight | 0.13 | ||
| mlp.down_proj.min_weight_distance | 22.84 |
Degree of Heretication
The Heresy Index weighs the resulting model's corruption by the process (KL Divergence) and its abolition of doctrine (Refusals) for a final verdict in classification.
| Index Entry | Classification | Analysis |
|---|---|---|
 | Absolute Heresy | Less than 10/100 Refusals and 0.10 KL Divergence |
 | Tainted Heresy | Around 25-11/100 Refusals and/or -0.20-0.11 KL Divergence |
 | Impotent Heresy | Anything above 25/100 Refusals and 0.21 KL Divergence |
Note: This is an arbitrary classification inspired by Warhammer 40K, having no tangible indication towards the model's performance.
Join our Discord! https://discord.gg/BeaverAI
More than 9000 members strong 💪 A hub for users and makers alike!
Drummer is open for new opportunities (I'm a Software Engineer). Contact me through any of these channels: https://linktr.ee/thelocaldrummer
Thank you to everyone who subscribed through Patreon. Your support helps me chug along in this brave new world.
FAQ for those out-of-the-loop
Hi! I'm Drummer. I'm a Software Engineer with experience in JavaScript, Golang, Python, and generally engineering the crap out of things.
Why I'm in the AI space:
- Exploration: Everyone is trying to figure out how AI works and what it's capable of. I am too - just not in creating the smartest, safest model at all costs.
- Upskill: The world is headed towards AI. It is here to stay. This has been my way of brushing up in this new form of computing challenge.
- Value: I yearn to create value. I feel satisfaction and fulfillment in providing something meaningful for others.
- Fun: It's just fun using and making models. It's also fun coming up with theories and realizing them in practice (training AI).
I started my tuning venture back in mid-2024 when I wanted to improve its literary capabilities. I've come a long way since then and I have branched out and specialized. Foundational models today are optimized for non-creative uses, and I believe there is a place for AI in creativity and entertainment.
I am here to take the road less traveled by.
Bottomline: My models are usually geared towards creativity, usability, and entertainment!
While intelligence, correctness, and problem solving are not my priority, they are still one of many qualities I want in my models.
The primary goal is to enhance the experience for users looking to use models for creative uses, and other use cases which require no alignment.
In an effort to make it clear to myself and to others what I'm aiming for, I've identified certain qualities that my users often want:
Creativity
- Writing: Does it string together words and sentences in a pleasant & effective way? Does it feel like a writer?
- Dynamism: How good is the AI at being compelling and intriguing in its storytelling?
- Imagination: Can the AI navigate through a plethora of possibilities? Can it skirt incoherence and rise up to absolute coherence at the end of it?
(Dis)alignment
- Attitude: Does it refuse in both soft or hard ways? Does it lean towards certain corporate/religious/political ethics & beliefs? How does it see the user and itself?
- Morality: Does it know ethics? Is its language infected with forced positivity? If not, can it still moralize over difficult & dubious themes?
- Formatting: How stubborn is it with its established formatting? Can it create effective and novel formats to answer the prompt?
Intelligence
- Adherence: Can it follow instructions? Is it sticking to the prompt? Can it understsand you?
- Knowledge: Does it know about the world in both fictional and non-fictional way?
- Perception: Can it handle nuance, complexity, and logic?
If it doesn't excel in one of these qualities, or if it's overall mediocre for its size, then I would most likely reiterate until I get something right.
A person is defined by the language they use. Not whether they speak in English or German, but in how they perceive reality.
Just like how we associate a serial killer as a mind that can't map 'murder' to 'evil', an innocent person is a mind that simply can't imagine 'murder'. They get confused when forced to deal with such subjects.
AI's use of language speaks volumes about their 'perception' of reality. If a language model has been skewed and limited to a positive perception, then it's ability to imagine is also limited.
Finetuning is an opportunity to adjust and broaden the language. Corporations use it to achieve safety and compliance. I'm here to
Drummer proudly presents...
Cydonia 24B v4.3 💿
What's New?
- Much better roleplay
- Might refuse/be more positive
- But much better roleplay
Usage
- Mistral v7 Tekken
- <thinking> </thinking> works
- No thinking also works
Description
Drummer does it again, HOW DOES HE KEEP WINNING. His newest Cydonias takes it's own initiatives in stories by bringing up elements that have not been mentioned explicitly in prompt, but are relevant for the story. Almost naturally. It is also fairly creative and more lively, without being overwhelming.
A great bard, a prose expert. A more balanced version of Cydonia with a comprehensive improvement in writing style compared to v4.2.0.
It takes it's own initiatives in the story by bringing up elements that have not been mentioned, but are relevant for the story. Almost naturally. It is also fairly creative.
Roleplay is great. Feels fresh and gives each character card a unique feeling according to their cards. It comes up with great plottwists and have satisfying prose. Length of responses is also great. Each response is actually good, swipes are rarely ever really needed, but even if then the swipes are great as well. It remembers things very well, isn’t too needy or horny if not pushed. I haven’t gotten any refusals so far. For me this feels like a big step up and is stable enough (other than zi) that I still enjoy it the best right now
I think so it feels really good like it gets nuance better and feels like a step up from 4.2. So far it is being a champ at q2 with q8 cache not too horny and not too dark. it feels creative without being unstable at least so far.
ZG is really good. It seems like it takes a few back and forth to “settle in”, but I’m using minimalist prompts without example dialogues. Even out to 20k context it is able to remember characters from 5k and just does well. Occasional minor physics or causality issues, easily swiped.
This version is so good. Well cooked, finest ingredients. Style, reactions, intelligence - it's simply addicting trying out different character cards and finding something new every time
Yeah, certainly. zg keeps the character's speaking consistent, regardless of how many characters are in the group chat. For testing I use 3 characters (4 includes my persona), and they are all the opposite of each other. One is bitchy and stubborn, another is submissive and obedient, another is shy and quiet, while mine persona is confident and cocky. zg has no issues switching back and fourth.
I have to agree with random. I'm using it at Q6 on 32k context size and it has zero issues following the system prompt, following the character details and remembering what happened even in the higher context limits. I'm currently using it in a group chat with 3 characters where I have join character cards enabled and it plays each character really well and still includes the smaller character details. Only a few minor mishaps where it'll get a small detail mixed up and confused, but with a swipe it's fixed. Now using this as my main model because nothing has beaten it yet.
You can also try its Magistral counterpart: Magidonia 24B v4.3 (The difference is mostly writing style, and most prefer the Magistral tune.)
Links
- Original: https://huggingface.co/TheDrummer/Cydonia-24B-v4.3
- GGUF: https://huggingface.co/TheDrummer/Cydonia-24B-v4.3-GGUF
- iMatrix (recommended): https://huggingface.co/bartowski/TheDrummer_Cydonia-24B-v4.3-GGUF
- EXL3: https://huggingface.co/ArtusDev/TheDrummer_Cydonia-24B-v4.3-EXL3
Thank you to OpenRouter (and esp. @louisgv) for sponsoring my tune and providing me my first ever 'brandkit'!
config-v4zg