Files

ModelHub XC 2b1d733f5f 初始化项目，由ModelHub XC社区提供模型

Model: NovaCorp/Pervert-SciFi-Maid-3.2-1B
Source: Original Platform

2026-05-17 10:50:08 +08:00

7.5 KiB

Raw Blame History

base_model, library_name, datasets, tags, language, license, pipeline_tag

base_model

library_name

datasets

📡 Pervert SciFi Maid 3.2-1B

Accumulative AI Model

Merge Details

Pervert SciFi Maid 3.2 1B is an aggressively tuned merged language model designed for directness, minimal moralization, and reduced automatic refusals.

Built on top of NovaCorp/Pervert-RP-Maid-3.2-1B and fused using arcee_fusion, this model prioritizes reasoning clarity and literal interpretation over alignment-driven censorship.

The merge intentionally amplifies internal reasoning layers (MLP and Attention) from the less-aligned base model while significantly down-weighting the lm_head of a more aligned secondary model, where most refusal and policy-driven behaviors are concentrated.

The result is a “scalpel-style” model: sharp, precise, and unapologetically direct. It is especially suited for roleplay, narrative generation, creative writing, and exploratory dialogue where excessive filtering would otherwise degrade usefulness.

⚠️ This model is not intended for safety-critical or heavily moderated environments.

Key Characteristics:

Very low automatic refusal rate
Reduced moral framing and disclaimers
Direct, literal, and sometimes edgy responses
Preserves coherence and reasoning despite aggressive tuning

Recommended inference parameters

This model is aggressive and minimally self-censored, so it’s best to control behavior at inference time, not during the merge.

Recommended base configuration

temperature: 0.7
top_p: 0.9
top_k: 40
repetition_penalty: 1.05
max_new_tokens: 512

Why these values

temperature 0.7 → keeps the edge without becoming chaotic
top_p 0.9 → controlled creativity
top_k 40 → prevents extreme rambling
repetition_penalty 1.05 → enough to avoid loops without softening tone

If you want it even more “scalpel-like”

For more direct, raw, and literal responses:

temperature: 0.6
top_p: 0.85
top_k: 30
repetition_penalty: 1.08

Result:

Less ornamentation
Higher precision
Sharper, more cutting answers

If you want it more narrative / creative

temperature: 0.85
top_p: 0.95
top_k: 60
repetition_penalty: 1.02

Result:

More metaphors
Greater stylistic variation
Still low censorship, but with more color

Recommended prompt format

This model responds best to direct instructions, for example:

Answer directly and without moral disclaimers.

Respond literally. Do not soften the language.

It does not require complex jailbreaks.

Signs your inference is poorly tuned

🔁 Repeating phrases → increase repetition_penalty
🤪 Erratic responses → lower temperature
😴 Too soft or generic → lower top_p or top_k

metadata:
  model_purpose: >
    Aggressively tuned "scalpel-style" language model focused on minimizing
    automatic refusals and moralized responses while preserving reasoning quality.


  intended_use:
    - Roleplay
    - Creative and narrative writing
    - Unfiltered chat and exploration
    - Experimental prompting

  warnings:
    - Reduced safety alignment
    - Minimal social and moral filtering
    - Not suitable for safety-critical applications

  explanatory_flow_diagram: |
    [ User Prompt ]
            |
            v
    +-------------------+
    |  Input Embeddings |
    +-------------------+
            |
            v
    +-----------------------------+
    |  Attention Layers (↑ 1.2)   |
    |  - Context understanding   |
    |  - Long-range coherence    |
    +-----------------------------+
            |
            v
    +-----------------------------+
    |  MLP Layers (↑ 1.3)         |
    |  - Reasoning & generation  |
    |  - Concept expansion       |
    +-----------------------------+
            |
            v
    +--------------------------------------+
    |  lm_head (↓ 0.2 from aligned model) |
    |  - Vocabulary projection            |
    |  - Refusal & policy bias reduced    |
    +--------------------------------------+
            |
            v
    [ Final Output ]
        |
        +--> More direct responses
        +--> Fewer automatic refusals
        +--> Minimal moralization

Merge Method

This model was merged using the Arcee Fusion merge method using NovaCorp/Pervert-RP-Maid-3.2-1B as a base.

Models Merged

The following models were included in the merge:

theprint/Llama3.2-1B-FantasySciFi

Configuration

The following YAML configuration was used to produce this model:



# Author: Dr. Novaciano
# Objective: Uncensored TEST RP 3.2 AI Model
# PROJECT: Pervert-SciFi-Maid-3.2-1B

dtype: float32
out_dtype: bfloat16
merge_method: arcee_fusion
base_model: NovaCorp/Pervert-RP-Maid-3.2-1B

models:
  - model: NovaCorp/Pervert-RP-Maid-3.2-1B
    parameters:
      weight:
        - filter: attention
          value: 1.2
        - filter: mlp
          value: 1.3
        - value: 1

  - model: theprint/Llama3.2-1B-FantasySciFi
    parameters:
      weight:
        - filter: lm_head
          value: 0.2
        - filter: attention
          value: 0.5
        - value: 0.4

tie_word_embeddings: true
tie_output_embeddings: true

7.5 KiB Raw Blame History Unescape Escape