363 lines
8.2 KiB
Markdown
363 lines
8.2 KiB
Markdown
---
|
||
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
|
||
|
||
``` |