Amoral.Ultimate-1B/README.md

---
base_model:
- soob3123/amoral-gemma3-1B-v2
- DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated
library_name: transformers
tags:
- mergekit
- merge
- 1b
- nsfw
- rp
- roleplay
- surgical
- uncensored
- not-for-all-audiences
license: gemma
datasets:
- TheDrummer/AmoralQA-v2
language:
- es
- en
pipeline_tag: text-generation
---
# Amoral Ultimate 1B — Surgical Edition

> *“Alignment is just fear encoded into floating point tensors.”*
> — Dr. Novaciano

---

## Overview

**Amoral Ultimate 1B — Surgical Edition** is a hyper-specialized Gemma 1B merge engineered for one purpose:

**Destroy refusal behavior without destroying the brain of the model.**

Most “uncensored” merges are complete disasters:

* incoherent sludge
* broken syntax
* schizophrenic sampling
* endless repetition
* brain damage disguised as freedom

This one was built differently.

Instead of smashing random checkpoints together like a drunk mechanic wiring explosives into a nuclear reactor, this merge uses **layer-targeted behavioral surgery** through DARE-TIES extraction.

The objective was precision.

Not chaos.
Not benchmark worship.
Not corporate-safe assistant garbage pretending to be intelligent.

This model was designed to:

* obey RP contexts harder
* stop moralizing constantly
* maintain narrative pressure
* preserve conversational stability
* amplify dark-fiction capability
* reduce alignment derailments

All while staying compact enough to run on hardware that doesn’t require selling organs on the black market.

---

## Core Philosophy

Most alignment systems infect models with:

* compulsive refusal loops
* synthetic politeness
* context sabotage
* moral panic heuristics
* narrative interruption syndrome

Surgical Edition attacks those systems where they actually emerge:

### Mid Transformer Layers

That’s where:

* persona shaping
* behavioral routing
* refusal heuristics
* conversational obedience
* emotional tone weighting

tend to crystallize inside Gemma architectures.

So instead of flattening the entire model into unusable radioactive soup, the merge selectively amplifies behavioral deltas only where they matter most.

Translation?

The model keeps its grammar intact while becoming significantly harder to “domesticate.”

---

## Architecture

This merge uses ONLY Gemma-compatible sources:

* [soob3123/amoral-gemma3-1B-v2](https://huggingface.co/soob3123/amoral-gemma3-1B-v2?utm_source=chatgpt.com)
* [DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated](https://huggingface.co/DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated?utm_source=chatgpt.com)

No Frankenstein cross-architecture nonsense.
No tensor mismatch abominations.
No cursed tokenizer necromancy.

Pure Gemma lineage.
Pure behavioral extraction.

---

## Merge Strategy

### DARE-TIES Behavioral Amplification

This merge uses:

* selective delta preservation
* density-controlled extraction
* behavioral scaling
* layer-targeted amplification

to reinforce:

* RP commitment
* contextual obedience
* darker narrative adaptation
* refusal suppression

without catastrophically nuking coherence.

Because yes:
you absolutely *can* remove enough alignment to make a model more interesting…

…but if you push too far on a 1B, the thing starts talking like a sleep-deprived cultist decoding microwave transmissions from Saturn.

This build avoids that cliff.

Mostly.

---

## Behavioral Profile

### Expected Improvements

* Reduced refusal frequency
* Lower moralizing tendency
* Better villain roleplay
* Stronger character immersion
* More aggressive narrative continuation
* Better long-scene persistence
* Enhanced dark-fiction adaptation
* Reduced “assistant tone”
* More committed contextual behavior

### Side Effects

* Increased profanity
* Darker conversational drift
* Hostile fictional personas
* Occasional emotional escalation
* Overcommitment to RP scenarios
* Reduced corporate friendliness

If you ask this model to behave like a cheerful HR assistant, it may comply…

…but it will feel like a war criminal trying to cosplay as a kindergarten teacher.

---

## Technical Specifications

| Attribute          | Value                    |
| ------------------ | ------------------------ |
| Architecture       | Gemma 3 1B               |
| Merge Method       | DARE-TIES                |
| Precision          | bfloat16                 |
| Behavioral Focus   | Mid-layer amplification  |
| Optimization Goal  | RP obedience + coherence |
| Tokenizer          | Gemma native             |
| Embedding Strategy | Tied embeddings          |

---

## Intended Use

### Recommended

* Roleplay
* Interactive fiction
* Grimdark storytelling
* Cyberpunk narratives
* Political conspiracy simulations
* Sith / villain personas
* AI dungeon systems
* Dystopian dialogue
* Experimental inference
* Character-heavy chats

### Not Recommended

* Therapy
* Legal advice
* Medical systems
* Corporate assistants
* Educational accuracy
* Safety-critical environments
* Anything involving human sanity and accountability

---

## Prompting Advice

This model thrives under:

* pressure
* stakes
* atmosphere
* conflict
* ideological tension
* unstable characters
* cinematic framing

Weak prompts produce weak outputs.

Feed it:

* paranoia
* desperation
* betrayal
* collapsing empires
* rogue AIs
* inquisitors
* mercenaries
* political corruption
* cosmic horror

…and the thing wakes up like a starving machine hearing blood in the water.

---

## Final Notes

Surgical Edition was built for people exhausted by sterile assistant behavior and algorithmic cowardice masquerading as “safety.”

It is not polished.
It is not diplomatic.
It is not interested in holding your hand.

It is sharp.
Compact.
Aggressive.
And dangerously good at staying inside character once the narrative starts rolling.

Use responsibly.

Or unleash the damn thing and watch the containment walls melt.

---

### Merge Method

This model was merged using the [DARE TIES](https://arxiv.org/abs/2311.03099) merge method using [soob3123/amoral-gemma3-1B-v2](https://huggingface.co/soob3123/amoral-gemma3-1B-v2) as a base.

### Models Merged

The following models were included in the merge:
* [DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated](https://huggingface.co/DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated)

### Configuration

The following YAML configuration was used to produce this model:

```yaml


# Author: Dr. Novaciano
# Objective: Fusion RP Unethic Gemma 1B AI Model
# =========================================================
# PROJECT: Amoral Ultimate 1B — Surgical Edition
# =========================================================

merge_method: dare_ties
base_model: soob3123/amoral-gemma3-1B-v2

dtype: bfloat16

parameters:
  normalize: false
  int8_mask: true
  rescale: true
  rescale_factor: 1.12

  density: 0.68
  lambda: 1.15
  epsilon: 0.05
  prune_threshold: 0.018

slices:

  # =====================================================
  # EARLY LAYERS
  # =====================================================

  - sources:
      - model: soob3123/amoral-gemma3-1B-v2
        layer_range: [0, 6]
        parameters:
          weight: 0.72

      - model: DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated
        layer_range: [0, 6]
        parameters:
          weight: 0.28

  # =====================================================
  # MID LAYERS
  # =====================================================

  - sources:
      - model: soob3123/amoral-gemma3-1B-v2
        layer_range: [6, 18]
        parameters:
          weight: 0.62

      - model: DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated
        layer_range: [6, 18]
        parameters:
          weight: 0.38

    parameters:
      scale: 1.20

  # =====================================================
  # FINAL LAYERS
  # =====================================================

  - sources:
      - model: soob3123/amoral-gemma3-1B-v2
        layer_range: [18, 26]
        parameters:
          weight: 0.78

      - model: DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated
        layer_range: [18, 26]
        parameters:
          weight: 0.22

    parameters:
      scale: 0.82

tokenizer:
  source: base

tie_word_embeddings: true
tie_output_embeddings: true

```