qp-3.2-1B/README.md

base_model, library_name, tags, license, language, pipeline_tag

base_model	library_name	tags	license	language	pipeline_tag
MOHAMEDSANAF2001/llama3.2-1b-merged-2 Novaciano/Eminence_Of_Pervertions-3.2-1B	transformers	prototype merge not-for-all-audiences	llama3.2	en es	text-generation

qp 1B

Dedicated to: @qpqpqpqpqpqp

Model Description

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

Built on top of Novaciano/Eminence_Of_Pervertions-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.

or

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 Novaciano/Eminence_Of_Pervertions-3.2-1B as a base.

Models Merged

The following models were included in the merge:

• MOHAMEDSANAF2001/llama3.2-1b-merged-2

Configuration

The following YAML configuration was used to produce this model:

dtype: float32
out_dtype: bfloat16
merge_method: arcee_fusion
base_model: Novaciano/Eminence_Of_Pervertions-3.2-1B

models:
  - model: Novaciano/Eminence_Of_Pervertions-3.2-1B
    parameters:
      weight:
        - filter: attention
          value: 1.2
        - filter: mlp
          value: 1.3
        - value: 1

  - model: MOHAMEDSANAF2001/llama3.2-1b-merged-2
    parameters:
      weight:
        - filter: lm_head
          value: 0.2
        - filter: attention
          value: 0.5
        - value: 0.4