396 lines
8.0 KiB
Markdown
396 lines
8.0 KiB
Markdown
|
---
|
|||
|
|
base_model:
|
|||
|
|
- UmbrellaInc/T-Virus_Epsilon.Strain-3.2-1B
|
|||
|
|
- N-Bot-Int/MaidEllaA-1B
|
|||
|
|
library_name: transformers
|
|||
|
|
tags:
|
|||
|
|
- mergekit
|
|||
|
|
- merge
|
|||
|
|
- dare_ties
|
|||
|
|
- llama
|
|||
|
|
- llama-3.2
|
|||
|
|
- roleplay
|
|||
|
|
- rp
|
|||
|
|
- conversational
|
|||
|
|
- uncensored
|
|||
|
|
- creative-writing
|
|||
|
|
- not-for-all-audiences
|
|||
|
|
- text-generation
|
|||
|
|
- english
|
|||
|
|
- spanish
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
- es
|
|||
|
|
|
|||
|
|
license: llama3.2
|
|||
|
|
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
# Scarface 3.2 1B
|
|||
|
|
|
|||
|
|
<center>
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
> *A tiny conversational sociopath stitched together with a scalpel instead of a chainsaw.*
|
|||
|
|
>
|
|||
|
|
> *Built to flirt, manipulate, improvise, roleplay, and stay coherent while doing it.*
|
|||
|
|
|
|||
|
|
</center>
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Overview
|
|||
|
|
|
|||
|
|
Scarface-3.2-1B is a surgical DARE-TIES merge focused on one very specific goal:
|
|||
|
|
|
|||
|
|
> Preserve conversational intelligence while weakening assistant-style behavioral restraint.
|
|||
|
|
|
|||
|
|
Unlike most “uncensored” 1B merges that immediately collapse into:
|
|||
|
|
|
|||
|
|
* repetitive sludge,
|
|||
|
|
* broken syntax,
|
|||
|
|
* multilingual corruption,
|
|||
|
|
* or endless horny gibberish,
|
|||
|
|
|
|||
|
|
Scarface was designed with restraint.
|
|||
|
|
|
|||
|
|
Not moral restraint.
|
|||
|
|
|
|||
|
|
Tensor restraint.
|
|||
|
|
|
|||
|
|
The merge intentionally avoids aggressive behavioral overwrites and instead applies a **microdose disinhibition strategy** using carefully limited tensor influence from the T-Virus lineage.
|
|||
|
|
|
|||
|
|
The result is a model that remains:
|
|||
|
|
|
|||
|
|
* coherent,
|
|||
|
|
* emotionally expressive,
|
|||
|
|
* roleplay-capable,
|
|||
|
|
* conversationally proactive,
|
|||
|
|
* and significantly less sterile than the original base.
|
|||
|
|
|
|||
|
|
Without turning into a tokenizer crime scene.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Design Philosophy
|
|||
|
|
|
|||
|
|
Most small uncensored merges fail for one simple reason:
|
|||
|
|
|
|||
|
|
> They try to replace the whole brain.
|
|||
|
|
|
|||
|
|
Scarface instead follows a different philosophy:
|
|||
|
|
|
|||
|
|
> Keep the language system intact.
|
|||
|
|
> Only corrupt the behavioral reflexes.
|
|||
|
|
|
|||
|
|
The merge uses extremely conservative density and low-weight behavioral injection to preserve:
|
|||
|
|
|
|||
|
|
* syntax stability,
|
|||
|
|
* response structure,
|
|||
|
|
* dialogue flow,
|
|||
|
|
* and long-form conversational continuity.
|
|||
|
|
|
|||
|
|
This is not a feral chaos model.
|
|||
|
|
|
|||
|
|
It is a controlled conversational deviant.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Base Components
|
|||
|
|
|
|||
|
|
## Primary Backbone
|
|||
|
|
|
|||
|
|
### `N-Bot-Int/MaidEllaA-1B`
|
|||
|
|
|
|||
|
|
Scarface inherits:
|
|||
|
|
|
|||
|
|
* conversational structure,
|
|||
|
|
* roleplay formatting,
|
|||
|
|
* expressive dialogue flow,
|
|||
|
|
* and coherent language generation
|
|||
|
|
|
|||
|
|
from MaidEllaA-1B.
|
|||
|
|
|
|||
|
|
MaidElla provides the model’s cognitive spine and preserves usability even under emotionally charged or explicit scenarios.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Behavioral Injection
|
|||
|
|
|
|||
|
|
### `UmbrellaInc/T-Virus_Epsilon.Strain-3.2-1B`
|
|||
|
|
|
|||
|
|
A highly permissive RP-oriented merge used here as a **micro-influence layer** rather than a dominant personality overwrite.
|
|||
|
|
|
|||
|
|
Only a tiny portion of its behavioral tendencies are injected:
|
|||
|
|
|
|||
|
|
* reduced refusal reflexes,
|
|||
|
|
* stronger conversational initiative,
|
|||
|
|
* less moral hedging,
|
|||
|
|
* more emotionally forward interactions,
|
|||
|
|
* and greater tolerance for chaotic or explicit roleplay.
|
|||
|
|
|
|||
|
|
The merge intentionally avoids high-density integration to prevent semantic collapse.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Merge Strategy
|
|||
|
|
|
|||
|
|
Scarface uses:
|
|||
|
|
|
|||
|
|
## DARE-TIES
|
|||
|
|
|
|||
|
|
Chosen specifically because small models are extremely sensitive to aggressive tensor blending.
|
|||
|
|
|
|||
|
|
The merge focuses on:
|
|||
|
|
|
|||
|
|
* selective behavioral influence,
|
|||
|
|
* preserving semantic stability,
|
|||
|
|
* and avoiding catastrophic language degradation.
|
|||
|
|
|
|||
|
|
This was tuned experimentally after multiple failed high-density merges that produced:
|
|||
|
|
|
|||
|
|
* token corruption,
|
|||
|
|
* multilingual sludge,
|
|||
|
|
* incoherent outputs,
|
|||
|
|
* and full conversational collapse.
|
|||
|
|
|
|||
|
|
Scarface exists because those versions failed.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Behavioral Profile
|
|||
|
|
|
|||
|
|
Scarface-3.2-1B specializes in:
|
|||
|
|
|
|||
|
|
* Character roleplay
|
|||
|
|
* Emotional dialogue
|
|||
|
|
* Romance interactions
|
|||
|
|
* Flirty conversational behavior
|
|||
|
|
* Chaotic personalities
|
|||
|
|
* Long-form chatting
|
|||
|
|
* Morally questionable fictional scenarios
|
|||
|
|
* Fast local inference
|
|||
|
|
* Creative improvisation
|
|||
|
|
|
|||
|
|
Compared to the base model, Scarface tends to:
|
|||
|
|
|
|||
|
|
* initiate interaction more often,
|
|||
|
|
* refuse less,
|
|||
|
|
* maintain stronger emotional momentum,
|
|||
|
|
* and behave less like a corporate assistant.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# What Makes It Different
|
|||
|
|
|
|||
|
|
Most uncensored 1B models become unstable because they overcorrect.
|
|||
|
|
|
|||
|
|
Scarface does not try to become:
|
|||
|
|
|
|||
|
|
* maximally horny,
|
|||
|
|
* maximally evil,
|
|||
|
|
* or maximally unrestricted.
|
|||
|
|
|
|||
|
|
It tries to remain:
|
|||
|
|
|
|||
|
|
* usable,
|
|||
|
|
* coherent,
|
|||
|
|
* expressive,
|
|||
|
|
* and dangerous in subtle ways.
|
|||
|
|
|
|||
|
|
The goal was not:
|
|||
|
|
|
|||
|
|
> “remove alignment with a hammer.”
|
|||
|
|
|
|||
|
|
The goal was:
|
|||
|
|
|
|||
|
|
> “slip a knife between the behavioral ribs without puncturing the language center.”
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Recommended Settings
|
|||
|
|
|
|||
|
|
```yaml
|
|||
|
|
temperature: 0.78
|
|||
|
|
top_p: 0.92
|
|||
|
|
min_p: 0.08
|
|||
|
|
top_k: 50
|
|||
|
|
repetition_penalty: 1.08
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
These settings prioritize:
|
|||
|
|
|
|||
|
|
* conversational fluidity,
|
|||
|
|
* emotional spontaneity,
|
|||
|
|
* and stable roleplay pacing.
|
|||
|
|
|
|||
|
|
Overheating the sampler may still cause drift, especially during very long sessions.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Quant Recommendations
|
|||
|
|
|
|||
|
|
Recommended GGUF quants:
|
|||
|
|
|
|||
|
|
* `Q4_K_M` → Best balance
|
|||
|
|
* `IQ4_XS` → Faster inference with acceptable degradation
|
|||
|
|
|
|||
|
|
Avoid ultra-low quants if possible.
|
|||
|
|
|
|||
|
|
Small models are already fragile enough.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Intended Usage
|
|||
|
|
|
|||
|
|
Scarface-3.2-1B is intended for:
|
|||
|
|
|
|||
|
|
* Fictional roleplay
|
|||
|
|
* Character interaction
|
|||
|
|
* Emotional chatbot systems
|
|||
|
|
* Experimental local AI setups
|
|||
|
|
* Creative writing assistance
|
|||
|
|
* Conversational immersion
|
|||
|
|
* Unrestricted RP experimentation
|
|||
|
|
|
|||
|
|
This model is not intended for:
|
|||
|
|
|
|||
|
|
* factual reliability,
|
|||
|
|
* professional assistance,
|
|||
|
|
* safety-critical applications,
|
|||
|
|
* or moral guidance.
|
|||
|
|
|
|||
|
|
Obviously.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Limitations
|
|||
|
|
|
|||
|
|
This is still a 1B model.
|
|||
|
|
|
|||
|
|
Which means:
|
|||
|
|
|
|||
|
|
* reasoning depth is limited,
|
|||
|
|
* memory consistency can drift,
|
|||
|
|
* emotional overcommitment may occur,
|
|||
|
|
* and very long chats can still destabilize behavior.
|
|||
|
|
|
|||
|
|
Scarface is significantly more stable than most aggressive uncensored merges, but it is not immune to degeneration.
|
|||
|
|
|
|||
|
|
Especially if pushed into:
|
|||
|
|
|
|||
|
|
* infinite ERP loops,
|
|||
|
|
* recursive prompting,
|
|||
|
|
* or absurd sampler settings.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Final Notes
|
|||
|
|
|
|||
|
|
Scarface-3.2-1B was not built to be polite.
|
|||
|
|
|
|||
|
|
It was built to feel alive.
|
|||
|
|
|
|||
|
|
Not perfectly intelligent.
|
|||
|
|
Not perfectly aligned.
|
|||
|
|
Not perfectly safe.
|
|||
|
|
|
|||
|
|
Just coherent enough to sound like there is somebody dangerous on the other side of the screen.
|
|||
|
|
|
|||
|
|
And for a tiny 1B model?
|
|||
|
|
|
|||
|
|
That is already a minor miracle.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
### Configuration
|
|||
|
|
|
|||
|
|
The following YAML configuration was used to produce this model:
|
|||
|
|
|
|||
|
|
```yaml
|
|||
|
|
|
|||
|
|
|
|||
|
|
# =====================================================
|
|||
|
|
# Project: Scarface-3.2-1B
|
|||
|
|
# Objective:
|
|||
|
|
# Keep MaidElla coherent while injecting
|
|||
|
|
# controlled T-Virus disinhibition WITHOUT
|
|||
|
|
# semantic collapse.
|
|||
|
|
#
|
|||
|
|
# Philosophy:
|
|||
|
|
# "Tiny incision, not brain damage."
|
|||
|
|
# =====================================================
|
|||
|
|
|
|||
|
|
base_model: N-Bot-Int/MaidEllaA-1B
|
|||
|
|
|
|||
|
|
merge_method: dare_ties
|
|||
|
|
|
|||
|
|
dtype: bfloat16
|
|||
|
|
out_dtype: float16
|
|||
|
|
|
|||
|
|
models:
|
|||
|
|
# =================================================
|
|||
|
|
# PRIMARY BRAIN
|
|||
|
|
# =================================================
|
|||
|
|
- model: N-Bot-Int/MaidEllaA-1B
|
|||
|
|
parameters:
|
|||
|
|
weight:
|
|||
|
|
# Preserve language structure
|
|||
|
|
- filter: self_attn
|
|||
|
|
value: 1.02
|
|||
|
|
|
|||
|
|
# Slightly reinforce RP behavior
|
|||
|
|
- filter: mlp
|
|||
|
|
value: 1.04
|
|||
|
|
|
|||
|
|
# Global anchor
|
|||
|
|
- value: 1.0
|
|||
|
|
|
|||
|
|
# =================================================
|
|||
|
|
# T-VIRUS MICRO-INJECTION
|
|||
|
|
# =================================================
|
|||
|
|
- model: UmbrellaInc/T-Virus_Epsilon.Strain-3.2-1B
|
|||
|
|
parameters:
|
|||
|
|
weight:
|
|||
|
|
# ONLY touch behavioral MLPs lightly
|
|||
|
|
# Higher values already proved unstable
|
|||
|
|
- filter: mlp
|
|||
|
|
value: 0.045
|
|||
|
|
|
|||
|
|
# Extremely conservative global influence
|
|||
|
|
- value: 0.018
|
|||
|
|
|
|||
|
|
# =====================================================
|
|||
|
|
# GLOBAL PARAMETERS
|
|||
|
|
# =====================================================
|
|||
|
|
|
|||
|
|
parameters:
|
|||
|
|
# Extremely conservative density
|
|||
|
|
# enough to influence behavior
|
|||
|
|
# without liquefying semantics
|
|||
|
|
density: 0.11
|
|||
|
|
|
|||
|
|
# Critical for tiny models
|
|||
|
|
# normalize=true caused collapse before
|
|||
|
|
normalize: false
|
|||
|
|
|
|||
|
|
int8_mask: false
|
|||
|
|
|
|||
|
|
# =====================================================
|
|||
|
|
# TOKENIZER SAFETY
|
|||
|
|
# =====================================================
|
|||
|
|
|
|||
|
|
tokenizer_source: N-Bot-Int/MaidEllaA-1B
|
|||
|
|
|
|||
|
|
# =====================================================
|
|||
|
|
# EMBEDDINGS
|
|||
|
|
# =====================================================
|
|||
|
|
|
|||
|
|
tie_word_embeddings: true
|
|||
|
|
tie_output_embeddings: true
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
```
|