georgesung/llama2_7b_chat_uncensored

Overview

Fine-tuned Llama-2 7B with an uncensored/unfiltered Wizard-Vicuna conversation dataset (originally from ehartford/wizard_vicuna_70k_unfiltered). Used QLoRA for fine-tuning. Trained for one epoch on a 24GB GPU (NVIDIA A10G) instance, took ~19 hours to train.

The version here is the fp16 HuggingFace model.

GGML & GPTQ versions

Thanks to TheBloke, he has created the GGML and GPTQ versions:

• https://huggingface.co./TheBloke/llama2_7b_chat_uncensored-GGML
• https://huggingface.co./TheBloke/llama2_7b_chat_uncensored-GPTQ

Running in Ollama

https://ollama.com/library/llama2-uncensored

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/llama2_7b_chat_uncensored.yaml

Fine-tuning guide

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

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	43.39
ARC (25-shot)	53.58
HellaSwag (10-shot)	78.66
MMLU (5-shot)	44.49
TruthfulQA (0-shot)	41.34
Winogrande (5-shot)	74.11
GSM8K (5-shot)	5.84
DROP (3-shot)	5.69