georgesung/llama3_8b_chat_uncensored

Overview

Fine-tuned Llama-3 8B with an uncensored/unfiltered Wizard-Vicuna conversation dataset. Used QLoRA for fine-tuning.

The model here includes the fp32 HuggingFace version, plus a quantized 4-bit q4_0 gguf version.

Prompt style

The model was trained with the following prompt style:

### HUMAN:
Hello

### RESPONSE:
Hi, how are you?

### HUMAN:
I'm fine.

### RESPONSE:
How can I help you?
...

Training code

Code used to train the model is available here.

To reproduce the results:

git clone https://github.com/georgesung/llm_qlora
cd llm_qlora
pip install -r requirements.txt
python train.py configs/llama3_8b_chat_uncensored.yaml

Fine-tuning guide

https://georgesung.github.io/ai/qlora-ift/

Ollama inference

First, install Ollama. Based on instructions here, run the following:

cd $MODEL_DIR_OF_CHOICE
wget https://huggingface.co/georgesung/llama3_8b_chat_uncensored/resolve/main/llama3_8b_chat_uncensored_q4_0.gguf

Create a file called llama3-uncensored.modelfile with the following:

FROM ./llama3_8b_chat_uncensored_q4_0.gguf
TEMPLATE """{{ .System }}

### HUMAN:
{{ .Prompt }}

### RESPONSE:
"""
PARAMETER stop "### HUMAN:"
PARAMETER stop "### RESPONSE:"

Then run:

ollama create llama3-uncensored -f llama3-uncensored.modelfile
ollama run llama3-uncensored