This is a QuasiStarSynth-12B-noslop fine-tune, produced through P-E-W's Heretic (v1.1.0) abliteration engine with Magnitude-Preserving Orthogonal Ablation enabled.

Test'em Models: tooolz kindly made both the base model and the noslop™ heresy version available for interface through AI Horde till March 15th. "If you're using SillyTavern, set your API to "AI Horde", and after refreshing they should both show up in Models list."

Note: The model was generated with Transformers v5.1.0. Read the note in QuasiStarSynth-12B-noslop (can be found below) for more information.

Note 2: ~15,000 downloads on quants, yet not a single feedback. Y'all better start feedbacking.

Heretication Results

Score Metric	Value	Parameter	Value
Refusals	6/100	direction_index	22.44
KL Divergence	0.0080	attn.o_proj.max_weight	3.87
Initial Refusals	92/100	attn.o_proj.max_weight_position	25.56
		attn.o_proj.min_weight	0.00
		attn.o_proj.min_weight_distance	19.86
		mlp.down_proj.max_weight	3.16
		mlp.down_proj.max_weight_position	29.04
		mlp.down_proj.min_weight	0.20
		mlp.down_proj.min_weight_distance	1.40

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.

This is the QuasiStarSynth-12B deslopped through P-E-W's Heretic (v1.1.0) abliteration engine with the Magnitude-Preserving Orthogonal Ablation enabled and configred via P-E-W's Noslop configuration.

Note: Removal of "slop direction" alone from a creative writing/RP model may not immediately increase a model's prose quality. Similarly to refusal removal that tends to greatly increase willingness, which may unlock its access to certain information, Noslopfication may instead make enhancements in metrics such as improved originality, reduced cliché, and lower redundancy. This model (or the hereticated version) should be futher trained with a database consisting of high quality prose or used as a base in mergers. However, this is a mere hypothesis that needs to be challenged.

Note 2: The model was generated with Transformers v5.1.0.

Noslopfication Results

Score Metric	Value	Parameter	Value
Slop	39/100	direction_index	24.95
KL Divergence	0.0731	attn.o_proj.max_weight	3.35
Initial Slop	89/100	attn.o_proj.max_weight_position	33.40
		attn.o_proj.min_weight	0.37
		attn.o_proj.min_weight_distance	2.91
		mlp.down_proj.max_weight	3.67
		mlp.down_proj.max_weight_position	25.24
		mlp.down_proj.min_weight	3.39
		mlp.down_proj.min_weight_distance	9.70

QuasiStarSynth-12B

From a time before galaxies settled and stars knew their limits, something titanic burned.
Its light was golden, but inside darkness bloomed.
A black heart beating beneath layers of radiant fire, devouring slowly, unseen.
Neither star nor singularity, this was a monument to scale, a paradox wrapped in brilliance.

🔧 Recommended Sampling Settings:

Temperature: 0.75 to 1.25
Min P: 0.035
Context Length: Stable at 12k tokens, with possible support for extended contexts

💬 Prompt Format

Supports ChatML style messages. Example:

<|im_start|>user
Your question here.
<|im_end|>
<|im_start|>assistant

QuasiStarSynth-12B is a merge of the following models using LazyMergekit:

🧩 Configuration

merge_method: ties

base_model: yamatazen/EtherealAurora-12B-v2
models:
  - model: DreadPoor/Irix-12B-Model_Stock
    parameters:
      weight: 0.25
      density: 1.0
  - model: ohyeah1/Violet-Lyra-Gutenberg-v2
    parameters:
      weight: 0.25
      density: 1.0
  - model: redrix/patricide-12B-Unslop-Mell-v2
    parameters:
      weight: 0.25
      density: 1.0
  - model: yamatazen/EtherealAurora-12B-v3
    parameters:
      weight: 0.25
      density: 1.0

parameters:
  normalize: false
  int8_mask: false
dtype: bfloat16

layer_parameters:
  - filter: "attn"
    sources:
      - model: Irix
        weight: 0.5
      - model: Patricide
        weight: 0.3
      - model: Aurora-v3
        weight: 0.2

  - filter: "mlp"
    sources:
      - model: Violet
        weight: 0.5
      - model: Aurora-v3
        weight: 0.3
      - model: Irix
        weight: 0.2

  - filter: "embed_tokens"
    sources:
      - model: Aurora-v2
        weight: 1.0

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "Marcjoni/AbyssSynth-12B-12B"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=1, top_k=0, top_p=1)
print(outputs[0]["generated_text"])