Overview

Fine-tuned Llama-2 13B with an uncensored/unfiltered Wizard-Vicuna conversation dataset ehartford/wizard_vicuna_70k_unfiltered. Used QLoRA for fine-tuning. Trained for one epoch on a two 24GB GPU (NVIDIA RTX 3090) instance, took ~26.5 hours to train.

{'train_runtime': 95229.7197, 'train_samples_per_second': 0.363, 'train_steps_per_second': 0.091, 'train_loss': 0.5828390517308127, 'epoch': 1.0}
100%|██████████████████████████████████████████████████████████████████████████████████████████████| 8649/8649 [26:27:09<00:00, 11.01s/it]
Training complete, adapter model saved in models//llama2_13b_chat_uncensored_adapter

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/Llama-2-13B-GGML
• https://huggingface.co/TheBloke/Llama-2-13B-GPTQ

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_13b_chat_uncensored.yaml

Fine-tuning guide

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