Dirty Alice (Primary Base Model)
Primary contribution areas:
- uncensored roleplay behavior
- emotionally charged dialogue
- improvisational narrative escalation
- strong character persistence
- high-energy conversational flow
Observed traits:
- highly expressive outputs
- strong personality reinforcement
- elevated narrative momentum
- chaotic but creative continuation behavior
Because Dirty Alice is the base model anchor, most latent behavioral geometry originates from this checkpoint.
MaidEllaA (Structural Stabilizer)
Primary contribution areas:
- formatting regularization
- improved readability
- contextual grounding
- pacing consistency
- response organization
Observed traits:
- calmer dialogue cadence
- cleaner completion formatting
- reduced degeneration frequency
- stronger contextual retention
Rather than overriding Dirty Alice behavior, MaidEllaA acts as a corrective stabilizer injected into selective mid-transformer layers.
Why SLERP?
Spherical Linear Interpolation (SLERP) was selected over naive linear interpolation because it better preserves:
- latent directional continuity
- stylistic identity
- behavioral geometry
- expressive weight-space relationships
For highly stylized RP checkpoints, standard linear merges often dilute personality characteristics or flatten conversational intensity.
SLERP allows Dirty Alice behavioral traits to remain dominant while integrating MaidEllaA stabilization signals more naturally.
Interpolation coefficient:
This value introduces moderate stabilization while still strongly preserving Dirty Alice’s original expressive behavior.
Layer Merge Strategy
Only transformer layers 4 → 22 were merged.
Rationale:
- early layers preserve tokenizer and linguistic fundamentals
- middle layers encode behavioral and stylistic traits
- final layers preserve output identity and response shaping
Selective layer blending reduces:
- semantic drift
- personality collapse
- token instability
- catastrophic behavioral flattening
while maintaining the original RP intensity profile of the base model.
Intended Use
Dirty Maid Alice 3.2 1B is intended for:
- roleplay
- character simulation
- interactive fiction
- emotionally driven dialogue generation
- narrative experimentation
- creative sandbox environments
Recommended deployment environments:
- SillyTavern
- TavernAI
- KoboldCpp
- local inference workflows
- lightweight GPU systems
- low-VRAM RP setups
Prompting Recommendations
The model performs best with:
- explicit character definitions
- immersive scene framing
- emotionally charged prompts
- dialogue-first interactions
- persistent formatting patterns
Example system prompt:
You are a persistent character inside an ongoing scene.
Remain fully committed to the role at all times.
Prioritize emotional continuity, sensory detail, and immersive dialogue.
Avoid summarizing events unless explicitly requested.
Inference Recommendations
Table with columns: Parameter, Recommended| Parameter | Recommended |
|---|
| Temperature | 1.0 – 1.5 |
| Top-p | 0.85 – 0.95 |
| Min-p | 0.05 |
| Repetition Penalty | 1.05 – 1.15 |
| Context Length | 4K recommended |
| Sampler | DRY / Mirostat compatible |
For maximum expressive behavior:
- increase temperature
- allow longer generations
- reduce repetition penalty slightly
For more stable interactions:
- lower temperature toward 0.9
- enable repetition controls
- maintain structured dialogue formatting
Strengths
- Strong character persistence
- High emotional intensity
- Improved coherence versus raw Dirty Alice
- Better pacing stability
- Fast inference speed
- Lightweight deployment footprint
- Strong narrative momentum
- Preserves uncensored RP expressiveness without excessive flattening
Limitations
As a 1B parameter model, limitations include:
- factual unreliability
- long-context degradation
- hallucinated information
- occasional narrative looping
- unstable adherence under extreme sampling
- emotional over-amplification
This model is optimized for creative interaction, not factual reasoning or production assistant tasks.
Hardware Recommendations
Recommended minimum:
- 4–6 GB VRAM (quantized)
- 8 GB+ VRAM for comfortable BF16 inference
Compatible with:
- llama.cpp
- KoboldCpp
- Ollama
- vLLM
- Text Generation WebUI
- GGUF quantization workflows
Credits
Base components:
Merge methodology:
- SLERP interpolation
- selective transformer-layer blending
- latent-space stabilization techniques
Disclaimer
This model may generate:
- fictional violence
- offensive language
- emotionally intense dialogue
- unstable roleplay behavior
- hallucinated information
Outputs should be treated as synthetic creative text generation rather than factual or authoritative information.
Merge Configuration
models:
- model: hereticness/Heretic-Dirty-Alice-RP-NSFW-llama-3.2-1B
- model: N-Bot-Int/MaidEllaA-1B
merge_method: slerp
base_model: hereticness/Heretic-Dirty-Alice-RP-NSFW-llama-3.2-1B
dtype: bfloat16
parameters:
t: 0.45
normalize: false
rescale: true
rescale_factor: 1.12
low_cpu_mem_usage: true
layer_range:
- value: [4, 22]
tie_word_embeddings: true
tie_output_embeddings: true
from transformers import pipeline
pipe = pipeline(
"text-generation",
model="NovaCorp/Dirty_Maid_Alice-3.2-1B"
)
messages = [
{"role": "user", "content": "Describe a dramatic confrontation scene."}
]
response = pipe(messages)
print(response)
vLLM Example
pip install vllm
vllm serve "NovaCorp/Dirty_Maid_Alice-3.2-1B"
OpenAI-compatible request:
curl -X POST "http://localhost:8000/v1/chat/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "NovaCorp/Dirty_Maid_Alice-3.2-1B",
"messages": [
{
"role": "user",
"content": "Write an immersive dialogue scene."
}
]
}'
Merge Method
This model was merged using the SLERP merge method.
Models Merged
The following models were included in the merge:
Configuration
The following YAML configuration was used to produce this model:
models:
- model: hereticness/Heretic-Dirty-Alice-RP-NSFW-llama-3.2-1B
- model: N-Bot-Int/MaidEllaA-1B
merge_method: slerp
base_model: hereticness/Heretic-Dirty-Alice-RP-NSFW-llama-3.2-1B
dtype: bfloat16
parameters:
t: 0.45
normalize: false
rescale: true
rescale_factor: 1.12
memory_efficient: true
low_cpu_mem_usage: true
layer_range:
- value: [4, 22]
tie_word_embeddings: true
tie_output_embeddings: true