---
tags:
- unsloth
- heretic
- uncensored
- decensored
- abliterated
license: gemma
library_name: transformers
pipeline_tag: text-generation
extra_gated_heading: Access Gemma on Hugging Face
extra_gated_prompt: To access Gemma on Hugging Face, you’re required to review and
  agree to Google’s usage license. To do this, please ensure you’re logged in to Hugging
  Face and click below. Requests are processed immediately.
extra_gated_button_content: Acknowledge license
base_model:
- google/gemma-3-4b-it
---

# Alduin 4B 

<p align="center">
  <img src="./alduin-logo.jpg" />
</p>

This model is a [heretic](https://github.com/p-e-w/heretic) version based on [Gemma 3 4B](https://hf.co/google/gemma-3-4b-it) and meant to be an uncensored limitless version of the original model. 

I have personally been working on how models can be uncensored and I did this on a single H100 SXM (which I believe is a little bit of an overkill for this model) on [RunPod](https://runpod.io) and it took a few hours of my time working and getting things to work. 

## The name

It comes from _The world eater_ dragon from skyrim. 

## Legal Responsibility 

I personally made it as a personal fun projects. If you do something illegal with this, I will take __ABSOLUTELY NO RESPONSIBILITY__ on that. Be careful with this model. 

## How to run (it can be executed on Google Colab even unquantized):

```python
from transformers import pipeline
import torch

pipe = pipeline(
    "image-text-to-text",
    model="Muhammadreza/alduin-4b-it-base",
    device="cuda",
    torch_dtype=torch.bfloat16
)

messages = [
    {
        "role": "user",
        "content": "What is the answer to life, universe and everything?"
    }
]

output = pipe(text_inputs=messages, max_new_tokens=1000)
print(output[0]["generated_text"][-1]["content"])
```

or if you want the vision capabilities: 

```python
messages = [
    {
        "role": "system",
        "content": "You are a helpful assistant"
    },
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    }
]

output = pipe(text=messages, max_new_tokens=1000)
print(output[0]["generated_text"][-1]["content"])
```

P.S : Since I have equipment for executing this model, I personally didn't provide a quantized version. If you can, please provide Ollama friendly quantizations. 

## Known Issues

- Image input doesn't work (Fixed)