TL;DR:
I created 3 bots with the same roleplay prompt with 3 LLMs: GPT 3.5, Mixtral, and Pygmalion, to test out the pros and cons of each LLM, as well as any fix there may be.
- GPT 3.5: Very decent base for beginners, the most stable and versatile one. A bit boring.
- Mixtral-8x7b-instruct: It is NOT built for roleplaying, and requires a lot of adjustment. However, once it is adjusted enough, the result is much more interesting than GPT 3.5.
- Pygmalion 13B: It is built for roleplaying. The best choice, but it is newly added to MyShell, and thus can be buggy and unstable.
Disclaimer:
I mainly create chatbots based on existing characters, so my record will focus more on chatbots that fall into the category of “famous characters”. However, I will still touch on other types of chatbots as well. Also, I don’t write prompts in the structured, hard coded way others do, so this guide will be more friendly to the beginners (like me), and to the experienced creators it will offer less guidance and more insight.
For this small experiment, I used the same prompt for three LLMs, which are GPT, Mixtral, and Pygmalion. I tried to recreate the character Xiao from the game Genshin Impact. Below are my opinions on the three LLMs, and at the end is an overview of the prompt I used.
