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

```yaml
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:

```yaml
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

```yaml
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`

---

```yaml
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](https://arcee.ai) merge method using [Novaciano/Eminence_Of_Pervertions-3.2-1B](https://huggingface.co/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](https://huggingface.co/MOHAMEDSANAF2001/llama3.2-1b-merged-2)

### Configuration

The following YAML configuration was used to produce this model:

```yaml
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

```