---
license: apache-2.0
base_model:
- mistralai/Mistral-7B-v0.1
tags:
- finetune
- mistral
widget:
  - text: "Avnas-7B-v1"
    output:
      url: https://cdn-uploads.huggingface.co/production/uploads/68e840caa318194c44ec2a04/eSPO6NFBEO0PpyMZ7oiOt.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 **Alpaca** template.
>

![avnas](https://cdn-uploads.huggingface.co/production/uploads/68e840caa318194c44ec2a04/eSPO6NFBEO0PpyMZ7oiOt.jpeg)

# Avnas 7B v1

This is my first model finetune so it may be a bit rough around the edges.

`mistralai/Mistral-7B-v0.1` was lightly trained on a custom dataset to uncensor the model and train it on Cthulhu mythos.

`v1.0 Training Steps: 100`

**The model appears to have no refusals and is fully uncensored, with no ablation needed.**

A version 2 might be released with larger dataset and longer cook time. This was mainly just proof of concept and it seems to work well.

**Avnas** was created using a custom finetuner kit I made called **PMPF** (Poor Man's Portable Finetuner). This allows finetuning with only 4-12GB VRAM. I could not get Axolotl, Unsloth, or any other tools working locally on Windows, so I made my own.

**Note:** Use `Alpaca` template to prevent errors. The dataset was specifically calibrated using `Alpaca` format due to issues with `ChatML` tokenizer.

`\n### Instruction:\n` `\n### Response:\n`

**Update:** The finetune EOS padding bug and safetensors were patched.

```py
<<<<<<
    # --- 4. Load Tokenizer ---
    tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, local_files_only=True)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "right"
======
    # --- 4. Load Tokenizer ---
    tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, local_files_only=True)
    
    # FIX: Use <unk> (ID 0) for padding instead of EOS (ID 2)
    # This prevents the model from learning to stop generating prematurely
    tokenizer.pad_token_id = 0  # unk_token_id for Mistral/Llama
    tokenizer.padding_side = "right"
>>>>>>
```