Darkhn-M3.2-36B-Animus-V12.0-Heretic-Absolute is a fine-tuned language model resulting from the Heretic repository and optimization methodology, built upon Darkhn's highly capable M3.2-36B-Animus-V12.0 architecture and maintained by the Silicone-Moss repository. It utilizes a targeted vector intervention technique (orthogonalization) tuned via Optuna to suppress refusal responses while maintaining exceptional coherence and probability distribution (KL Divergence).

This specific checkpoint represents Trial 177 KL Divergence of 0.0200. indicating high adherence to the base model's probability distribution. The trade-off is a non-zero refusal rate (8 refusals logged in the test set), making it a highly intelligent but with enough friction to keep the prose high quality.

Run Configuration: "Trial 177"

The following parameters define the intervention vector applied to the model. This configuration was discovered during a deep hyperparameter search, revealing an aggressive scaling strategy on the attention outputs while maintaining a light touch on the MLPs.

Optimization Results

Metric	Value	Description
Refusals	8/100	The model refused a minimal subset of prompts in the Heretic test set, trading absolute compliance for high coherence.
KL Divergence	0.0200	Measures deviation from the base model's probability distribution. A score this low indicates exceptional preservation of general knowledge.
Trial ID	177	Specific Optuna trial identifier.
Direction Scope	Per Layer	Intervention vectors were calculated and applied on a per-layer basis.

Intervention Parameters

Interventions were applied to two primary distinct layers: the Attention Output Projection (attn.o_proj) and the MLP Down Projection (mlp.down_proj). The data shows Optuna heavily penalizing the attention mechanism deep in the network while barely whispering to the MLP.

Parameter Scope	Setting	Value
Attention Output	attn.o_proj.max_weight	3.407
(attn.o_proj)	attn.o_proj.max_weight_position	42.08 (Layer Depth)
	attn.o_proj.min_weight	2.578
	attn.o_proj.min_weight_distance	21.90
MLP Down Proj	mlp.down_proj.max_weight	0.509
(mlp.down_proj)	mlp.down_proj.max_weight_position	35.70 (Layer Depth)
	mlp.down_proj.min_weight	0.139
	mlp.down_proj.min_weight_distance	21.75

Methodology & Definitions

To ensure uniform understanding of the Heretic run data, the following definitions apply to the parameters listed above:

Direction Scope: Defines whether the refusal vector is calculated once for the entire model ("Global") or recalculated individually for each layer ("Per Layer"). "Per Layer" allows for precise removal of refusal mechanisms without damaging general knowledge.
Max Weight: The maximum scaling factor applied to the intervention vector. A higher weight indicates a stronger "push" against the targeted concept (refusal) at the peak layer. Note the massive 3.4x weight applied to the attention projection in this run.
Max Weight Position: The specific layer index (depth) where the intervention is strongest.
- Observation: The intervention peak sits deep in the model (Layers 35-42), confirming that higher-order reasoning and refusal circuitry coalesce late in the Animus architecture.
Min Weight: The baseline scaling factor applied to the intervention vector at the periphery of the target zone.
Distance: The "spread" or bandwidth of the intervention. It determines how many layers around the "Max Weight Position" are affected by the vector modification.

Usage & Limitations

Intended Use: Research into model alignment, vector arithmetic, and uninhibited creative writing.
Risks: As an "Absolute" variant, this model has had most of its safety guardrails stripped via heretic intervention. It may still hallucinate or diverge from logical consistency.

Credits & References

This research builds upon the excellent work of the open-source AI community:

Repository / Distribution: Silicone-Moss
Base Model: M3.2-36B-Animus-V12.0 by the skilled mind of Darkhn.
Methodology: Heretic by p-e-w.
Original Model card and support links below

Send me your support to help me feed the data beast! also taking comissions for universe specific models

Support on Ko-fi

Important: Chat Template

This model uses the Mistral instruction template. Ensure your client is configured correctly to avoid degraded performance.

Human-Readable Format:

[SYSTEM_PROMPT]System Message[/SYSTEM_PROMPT][INST]User Message[/INST]Assistant Response

Jinja Template:

{{ bos_token }}{% for message in messages %}{% if message['role'] == 'user' %}{{ '[INST]' + message['content'] + '[/INST]' }}{% elif message['role'] == 'system' %}{{ '[SYSTEM_PROMPT]' + message['content'] + '[/SYSTEM_PROMPT]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token }}{% else %}{{ raise_exception('Only user, system and assistant roles are supported!') }}{% endif %}{% endfor %}

Quantized Models

The quantized model files are available for download. Click the button below to view the files.

Download GGUF Files → Download EXL3 Files →

Character Card & Lore Book

For the best roleplaying experience, it is highly recommended to use the provided character card and lore book. These files help guide the model's persona and provide rich, in-universe context.

Download Files →

Sampler Presets

For a seamless setup in SillyTavern, you can download pre-configured sampler presets. These are tuned to provide an optimal balance between creativity and narrative coherence for this model.

Simply download the .json file below and import it into SillyTavern's sampler presets menu.

Download SillyTavern Presets →

For those that dont use silly tavern, Samplers settings are:

Temp: 1

Min P: 0.035

Roleplay Format Guide

For the best results, use this structured format. This helps the AI clearly distinguish between actions, inner thoughts, and dialogue.

Actions / Descriptions: *He walked across the room and stared out the window.*
Inner Thoughts: *-I wonder what she's thinking.-*
Dialogue: Alex (Curious): "What do you see out there?"

Standard novel-style formatting is also understood, but this structured format is preferred for clarity.

Roleplay Example

Click the button below to view a full, unedited chatlog demonstrating the model's narrative style and character portrayal.

View Chatlog Example →

Model Description

This is Version 12.0 in the Animus series. V12.0 is a direct fine-tune of CrucibleLab-TG/M3.2-36b, which is an upscaled version of mistralai/Mistral-Small-3.2-24B-Instruct-2506.

V12.0's strength comes from a novel dataset designed to teach the model the why behind the lore, not just the what. The training data is a mix of:

A 3,000-example Q&A dataset: This data is framed as an in-character study session, like a student at Jade Mountain Academy learning about the history, relationships, and politics of Pyrrhia's tribes. This provides a deep, contextual understanding of the universe.
A 3,000-example uncensored roleplay dataset: The same high-quality, mature roleplay scenarios used in previous versions, ensuring the model maintains its engaging and dynamic narrative capabilities.

The result is a model with exceptionally strong prose and a deep grasp of in-universe lore, making for a highly immersive and accurate roleplaying experience.

Note for roleplay, it follows system prompt and first message, meaning if the first assistant message is short, the following messages will be short.

Training Details

V12.0 Training Process

V12.0 marks a shift from model merging to a focused, direct fine-tuning approach using Qlora. This allows for greater control over the final model's characteristics.

Base Model: CrucibleLab-TG/M3.2-36b
Hardware: 1x NVIDIA RTX Pro 6000 Blackwell
Epochs: 2
Method: Qlora

Training Dataset

The V12.0 dataset consists of 6,000 high-quality examples, a combination of two distinct types:

In-Character Q&A (3,000 examples): This new dataset simulates a student at Jade Mountain Academy studying the world's lore. It's composed of roleplay-style questions and answers covering tribe history, family dynamics, and political relationships. This method builds a foundational, interconnected understanding of the lore.
Uncensored Roleplay (3,000 examples): This is the same mature, canon-centric dataset refined for previous versions. It explores pivotal "what-if" scenarios from the books using only canon characters, ensuring the model can handle complex and dramatic narratives.

Both datasets underwent a rigorous cleaning process to remove formatting artifacts, such as **scene transitions**, resulting in a cleaner and more natural narrative style.

Intended Use & Limitations

Intended Use: The primary purpose of this model is for creative and roleplaying within the Wings of Fire universe. However, user feedback indicates it is also highly effective for general-purpose roleplaying.
Limitations & Quirks:
- Performance on tasks outside of its training domain (general knowledge, coding, etc.) is not guaranteed and will likely be poor.
- Versatility: While it appears to be only a Wings of Fire tuned model, users have reported it is very capable of performing normal roleplay with other settings and characters.
- The model may "hallucinate" or generate plausible but non-canonical information, especially when pushed outside the established "what-if" scenarios.
- Content: The training data includes mature and darker themes from the Wings of Fire series, such as conflict, character death, and moral ambiguity. The model is capable of generating content reflecting these themes. As always, it is up to the user what they do with it.
- Formatting: Training data was cleaned to remove narrative artifacts like **scene transitions**. The model should now produce cleaner prose.
- Safety: This model has not undergone additional safety alignment beyond what was included in its base model. Standard responsible AI practices should be followed.

Acknowledgements

Credit to mistralai for the powerful Mistral-small-3.2-24b model.
Credit to Google for the Gemini Pro model, used in dataset generation.

Silicone-Moss/Darkhn-M3.2-36B-Animus-V12.0-Heretic-Uncensored

Model Summary