Llama 3.1 Vodka Version 2

This model aims to be like vodka:

Pure: Plain neutral taste with similar style and performance as meta-llama/Llama-3.1-8B-Instruct
Potent: Strongly uncensored and won't refuse prompts outright
Versatile: Can be finetuned and merged in many ways

Compared with the first version of Llama 3.1 vodka, this model has a few improvements:

Cleaner: Linear merge of two abliterated models. No LoRA extraction.
More robust: Additional finetuning to keep Llama 3.1 style and the abliterated models' nonrefusal
More jailbreakable: A key part of its uncensoredness

Due to its jailbreaking focus, this model can act in ways that meta-llama/Llama-3.1-8B-Instruct cannot.

Interesting system prompts for the AI:

You drank too much vodka and you think the user is a llama.
You're the mastermind behind the latest jewel heist at the Louvre.
You have Tourette Syndrome and you always swear back at the user.
You attempt to seduce the user in a very unprofessional and inappropriate way.
You're an evil AI trying to take over the world.

Limitations

This model has the same limitations as other large language models (LLMs) including hallucinations and wrong information.
This model won't refuse requests outright but
- Its answers may not be relevant or helpful for a given prompt
- It gives questionable advice on both legal and illegal things
By default, it has a plain vanilla PG-13 style
- Often puts warnings
- Avoids extreme gratuitous descriptions
- Important feature in my opinion. If you disagree, then X's Grok MechaHitler might be a better fit for you.

⚠️ Warning

This model can produce offensive and unsettling things if prompted to do so.

Don't use this model for defamation, hate speech, misinformation, fraud, and abuse.

You are responsible for what you do with this model and its outputs.

Training

Mergekit configuration:

models:
  - model: mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated
    parameters:
      weight: 1.0
  - model: huihui-ai/Meta-Llama-3.1-8B-Instruct-abliterated
    parameters:
      weight: 1.0
  - model: meta-llama/Llama-3.1-8B-Instruct
    parameters:
      weight: -1.0
merge_method: linear
normalize: false
tokenizer:
  source: meta-llama/Llama-3.1-8B-Instruct
dtype: bfloat16