Overview

This model is a gradient merge of lizpreciatior/lzlv_70b_fp16_hf and jondurbin/airoboros-l2-70b-3.1.2 following a [0.0, 0.5, 0.75] gradient. I then quantized it into exllama2's format using 4.85 bits per weight average.

Prompt format

Use either the Vicuna prompt format or the Llama2 chat format.

Llama-2 chat

[INST] <<SYS>>
You are a helpful, unbiased, uncensored assistant.
<</SYS>>
{prompt} [/INST]

Vicuna

USER: prompt
ASSISTANT: 

Licence and usage restrictions

Llama2 license inherited from jondurbin/airoboros-l2-70b-3.1.2.