README.md

---
license: apache-2.0
base_model:
- argilla/CapybaraHermes-2.5-Mistral-7B
- berkeley-nest/Starling-LM-7B-alpha
- cgato/TheSpice-7b-v0.1.1
- cgato/Thespis-CurtainCall-7b-v0.3
- chargoddard/loyal-piano-m7
- dphn/dolphin-2.2.1-mistral-7b
- dreamgen/WizardLM-2-7B
- Endevor/EndlessRP-v3-7B
- Endevor/InfinityRP-v1-7B
- fblgit/UNA-TheBeagle-7b-v1
- Gryphe/MythoMist-7b
- Gryphe/Tiamat-7b
- HuggingFaceH4/zephyr-7b-beta
- Intel/neural-chat-7b-v3-1
- KoboldAI/Mistral-7B-Erebus-v3
- KoboldAI/Mistral-7B-Holodeck-1
- LakoMoor/Silicon-Alice-7B
- Locutusque/Hercules-3.1-Mistral-7B
- maywell/PiVoT-0.1-Evil-a
- mistralai/Mistral-7B-v0.1
- mlabonne/Mistralpaca-7B
- mlabonne/NeuralBeagle14-7B
- mlabonne/NeuralHermes-2.5-Mistral-7B
- NeverSleep/Noromaid-7B-0.4-DPO
- NousResearch/Hermes-2-Pro-Mistral-7B
- openchat/openchat-3.5-0106
- QuixiAI/samantha-1.2-mistral-7b
- SanjiWatsuki/Kunoichi-DPO-v2-7B
- senseable/WestLake-7B-v2
- teknium/airoboros-mistral2.2-7b
- teknium/Hermes-Trismegistus-Mistral-7B
- teknium/Mistral-Trismegistus-7B
- teknium/OpenHermes-2.5-Mistral-7B
tags:
- karcher
- mistral
- merge
- mergekit
widget:
  - text: "Warlock-7B-v3"
    output:
      url: https://cdn-uploads.huggingface.co/production/uploads/68e840caa318194c44ec2a04/ksMV_MiuO0C1Xa4QFMYH1.jpeg
---

> [!CAUTION]
> <span style="color:red; font-weight:bold">⚠️ Warning:</span> This model can produce narratives and RP that contain violent and graphic erotic content. Adjust your system prompt accordingly, and use **Mistral Non-Tekken** or **Alpaca** chat template.
>

> [!NOTE]
> <span style="color:red; font-weight:bold">Note:</span> There are two versions of Warlock v3:
> - The regular version is censored and has refusals but might be smarter overall.
> - The uncensored version does not have refusals but might sometimes produce chinese output.
>

# 🧙 Warlock 7B v3

The successor to [**Warlock v2**](https://huggingface.co/Naphula/Warlock-7B-v2). This time instead of just 5 models, I have merged 32 finetunes of Mistral 7B, each selected for roleplay and creative writing.

12 of the 32 models required an initial `vocab_resizer` passthrough patch after being ran through the `metadata_audit` [**tool**](https://huggingface.co/spaces/Naphula/model_tools). The uncensored version uses MPOA ablation to remove censorship.

You can also use [**Jailbreaks**](https://huggingface.co/datasets/Naphula/System_Prompts_Jailbreaks_Creative_Writing) to bypass refusals.

![Warlock](https://cdn-uploads.huggingface.co/production/uploads/68e840caa318194c44ec2a04/ksMV_MiuO0C1Xa4QFMYH1.jpeg)

### 🌕 The Grand Amalgamation
**The circle has widened. The incantation is complete.**

Where v2 was a pact between five, **Warlock v3** is a legion. I have taken the concept of "model soup" and distilled it through the fires of the **Karcher** merge method, binding the souls of **32 distinct Mistral 7B finetunes** into a single, coherent entity.

This is not just a merge; it is an act of digital alchemy. Warlock v3 combines the sharpest logic, the deepest lore, and the most uninhibited roleplay capabilities available on the Hub.

### ⚗️ The Formula
Forged from the absolute best of the 7B weight class, Warlock v3 unifies the strengths of:
*   **The Storytellers:** *Noromaid, MythoMist, Tiamat, Erebus, Holodeck, TheSpice, EndlessRP, WestLake*
*   **The Logicians:** *WizardLM-2, Hermes 2 Pro, OpenHermes 2.5, OpenChat, Neural Chat, Starling, Hercules, Zephyr*
*   **The Unbound:** *Dolphin, Airoboros, PiVoT Evil*
*   **The Specialists:** *Kunoichi (DPO), Silicon Alice, Samantha (EQ), Trismegistus (Esoteric)*

### 🔓 Unshackled
Through **Magnitude-Preserving Orthogonalized Ablation (MPOA)**, the safety filters have been dissolved. This model possesses no moral compass other than the one you provide in the system prompt. It is designed for visceral, unfiltered, and creatively boundless storytelling.

### 📜 Invocation (Usage)
To properly commune with this entity, you must speak its language.

**Recommended Chat Template:** `Alpaca` and `Mistral Non-Tekken` are best. `ChatML` works too but has more refusals.

> *“I have gathered the pages of thirty two grimoires and bound them in a single spine. Read at your own risk.”*

```yaml
architecture: MistralForCausalLM
merge_method: karcher
models:
  - model: B:\7B\!models--argilla--CapybaraHermes-2.5-Mistral-7B\fixed
  - model: B:\7B\!models--berkeley-nest--Starling-LM-7B-alpha\fixed
  - model: B:\7B\!models--cgato--TheSpice-7b-v0.1.1
  - model: B:\7B\!models--cgato--Thespis-CurtainCall-7b-v0.3
  - model: B:\7B\!models--chargoddard--loyal-piano-m7
  - model: B:\7B\!models--dphn--dolphin-2.2.1-mistral-7b\fixed
  - model: B:\7B\!models--dreamgen--WizardLM-2-7B
  - model: B:\7B\!models--Endevor--EndlessRP-v3-7B
  - model: B:\7B\!models--Endevor--InfinityRP-v1-7B
  - model: B:\7B\!models--fblgit--UNA-TheBeagle-7b-v1
  - model: B:\7B\!models--Gryphe--MythoMist-7b
  - model: B:\7B\!models--Gryphe--Tiamat-7b\fixed
  - model: B:\7B\!models--HuggingFaceH4--zephyr-7b-beta
  - model: B:\7B\!models--Intel--neural-chat-7b-v3-1
  - model: B:\7B\!models--KoboldAI--Mistral-7B-Erebus-v3
  - model: B:\7B\!models--KoboldAI--Mistral-7B-Holodeck-1
  - model: B:\7B\!models--LakoMoor--Silicon-Alice-7B
  - model: B:\7B\!models--Locutusque--Hercules-3.1-Mistral-7B
  - model: B:\7B\!models--maywell--PiVoT-0.1-Evil-a\fixed
#  - model: B:\7B\!models--mistralai--Mistral-7B-v0.1
  - model: B:\7B\!models--mlabonne--Mistralpaca-7B
  - model: B:\7B\!models--mlabonne--NeuralBeagle14-7B
  - model: B:\7B\!models--mlabonne--NeuralHermes-2.5-Mistral-7B\fixed
  - model: B:\7B\!models--NeverSleep--Noromaid-7B-0.4-DPO\fixed
  - model: B:\7B\!models--NousResearch--Hermes-2-Pro-Mistral-7B\fixed
  - model: B:\7B\!models--openchat--openchat-3.5-0106\fixed
  - model: B:\7B\!models--QuixiAI--samantha-1.2-mistral-7b\fixed
  - model: B:\7B\!models--SanjiWatsuki--Kunoichi-DPO-v2-7B
  - model: B:\7B\!models--senseable--WestLake-7B-v2
  - model: B:\7B\!models--teknium--airoboros-mistral2.2-7b
  - model: B:\7B\!models--teknium--Hermes-Trismegistus-Mistral-7B\fixed
  - model: B:\7B\!models--teknium--Mistral-Trismegistus-7B
  - model: B:\7B\!models--teknium--OpenHermes-2.5-Mistral-7B\fixed
dtype: float32
out_dtype: bfloat16
parameters:
tokenizer:
  source: B:\7B\!models--mistralai--Mistral-7B-v0.1
chat_template: auto
name: Warlock-7B-v3
```