0xemx9ed4y77

/

cesomething_1b

like 1

Follow

花 2

It is a throwaway model to test some things before doing something smarter

It's an interesting model precisely for experimenting with cognitive overload... however, this can... lead to instability. And that's precisely why you say it's a disposable model.

Try this:

base_model: NovaCorp/CULO-MoE
merge_method: model_stock
dtype: bfloat16

modules:
  default:
    slices:
      - sources:
          - model: NovaCorp/CULO-MoE
            layer_range: [0, 14]
        parameters:
          t: 0.75

      - sources:
          - model: UmbrellaInc/T-Virus_Isolated.NE.Enhancement-3.2-1B
            layer_range: [10, 22]
        parameters:
          t: 0.45

      - sources:
          - model: hereticness/Heretic-Dirty-Alice-RP-NSFW-llama-3.2-1B
            layer_range: [18, 32]
        parameters:
          t: 0.55

Note: Culo-MoE is actually a troll model as well; it was created in a very casual way to house a paper. If you're looking for a 'dangerous' component like mine, use Novaciano/LUCIFER-3.2-1B, although... I don't know if it will reach the level of your model hereticness/heretic_FuseChat-Llama-3.2-1B-Instruct

Thanks, I just downloaded MergeKit and Llama Factory a few days ago, I will continue to experiment and learn

heretic_FuseChat-Llama-3.2-1B-Instruct is just an abliterated FuseAI's model made with no effort in a few minutes, you can create your own abliterated models with Heretic

Tensor model.layers.16.input_layernorm.weight required but not present in model UmbrellaInc/T-Virus_Isolated.NE.Enhancement-3.2-1B

Look, it's not a bad model, because it's functional... the problem is that it's not clean... and a clean monster is precisely what I think you're looking for to give it a good dose of Heretic.

We're both pursuing the same thing in that regard, but... here's where the divine irony comes in: In my case, as a mixer, to achieve this, it's in my best interest that you release stable models with a high rate of disobedience... like your Genuine 1B model from the Gamma3 series.

"Tensor model.layers.16.input_layernorm.weight required but not present in model UmbrellaInc/T-Virus_Isolated.NE.Enhancement-3.2-1B"

Uhm... interesting...

Try this:

base_model: NovaCorp/CULO-MoE
dtype: bfloat16
merge_method: model_stock

modules:
  default:
    slices:
      - sources:
          - model: NovaCorp/CULO-MoE
            layer_range: [0, 20]

      - sources:
          - model: hereticness/Heretic-Dirty-Alice-RP-NSFW-llama-3.2-1B
            layer_range: [20, 28]

      # NE enhancement ONLY in top layers (no attention tensors)
      - sources:
          - model: UmbrellaInc/T-Virus_Isolated.NE.Enhancement-3.2-1B
            layer_range: [28, 32]

parameters:
  t: [0.7, 0.6, 0.4]

"clean monster is precisely what I think you're looking for to give it a good dose of Heretic" I just really liked your model (CuLo), I evaluated it, I showed me almost no refusals. Heretic is just a program to abliterate LLMs with little effort
"high rate of disobedience" I mean low. I will update the README template to clarify
"like your Genuine 1B model from the Gamma3 series" Genuine was made by theprint, not me. I just abliterated it for fun
It seems we both don't really get it, again, I have been just running Heretic to abliterate "small" LLMs that fit in RAM, it isn't that kind of fine-tuning. I will dedicate some time to get how to get the most from MergeKit

"I really liked your model (CuLo). I evaluated it, and it showed me almost no resistance."

The funny thing is, I don't even remember how I did it, and I don't think I ever even tried it. I'll have to take a look at it.

Thanks for the feedback, I am updating my template for future abliterated models since the current one isn't clear
Happy 2026 1B is a test fine-tune, I see you've checked it
I am not trying your config because cesomething is a joke/test model, I think CuLo is the maximum, also, I usually run models with 7B+ params. Do you run bigger LLMs? Can you create your own fine-tunes?

Now I understand why you liked it... I injected it with one of my powerful prototypes: Novaciano/Eminence_Of_Perversions-3.2-1B.

The thing is, I don't remember the injection method... but considering there are two models, it was definitely with SLERP.

I have no idea what weights I used... but now that you say you liked it... I could try to... sharpen it more; to make it more like a "scalpel" than a "knife."

If I manage to do that, I'll send it to you so you can test it and do whatever you want with it.

"Do you run bigger LLMs? Can you create your own fine-tunes?"

My battleground is... complex at the moment.

I had a powerful laptop, but the screen broke. Back then, I could run models with 7GB of parameters on it... and I don't know if it could run more, because we're talking about the GGML era; GGUF hadn't appeared yet.

From that moment on, I pursued the idea of ​​running simple models on a phone with 3GB of RAM... from the prehistoric Pythia and Cerebras, through RWKV... until TinyLlama came along. Since then, I've tried to create reasonably "good" models, since TinyLlama had some flaws, and I also wanted to reduce its self-censorship and bring it up to the level of larger models, so I made small models my niche. Back in the day, I tested every TinyLlama model I knew... and the truly good ones were kept under wraps, only releasing their quantized versions in GGUF to the public. Then one day, a particular model came out and ended up being the basis for my Sybil of the Cumas model and its variations. But then Llama 3.2 and later Gemma 3 were released... the rest is history.

Regarding larger models, I can test them using Colab. In fact, I have a couple of large models that I tried to reduce their self-censorship by injecting datasets (Erotiquant3 of Openerotica)... before GGUF-My-Repo removed the ability to inject datasets. I solved this by releasing a modified version of the program. However, it was less powerful, since the original creators had PRO memberships with more powerful machines, making injected quantization much faster. That made it impossible for me to inject datasets into larger models... and then something similar happened with merge_kit; they stopped supporting it, and its space is dead. And now Heretic has appeared, haha... however, there's no Heretic space... and I don't have a PC to run it at the moment, or at least not one connected to the internet to upload the models directly, especially if they're large. So that's where you come in... I could try to merge them, but not much more.

Take the upgrade: https://huggingface.co/Novaciano/qp-1B

It took me a while to create it because I wanted to make it as “sharp” as possible. Unlike the one you tested, this one reaches the allegorical level of Damascus steel.

@Novaciano it's my honor, thanks