This project is a work in progress - the foundations are not yet complete.
This project serves as a way for me to learn about the nature of the responses of LLMs when they are given an objective/scenario. In this case, our LLM is the brain of an autonomous Minecraft bot.
I find the way that LLMs act and respond very fascinating. More specifically, how well they process different kinds of data and when they choose to carry out certain tasks in an agentic loop. I am also trying to find out how changing the formatting of a prompt and the data given to the LLM will affect how effectively it can interpret and respond to the input.
- LLM: Qwen3.5 via Ollama running locally
- Minecraft client: Mineflayer
Must have a .env file in the root directory with the following fields:
- MINECRAFT_PORT
- MINECRAFT_VERSION
The current prompt is causing the LLM to hallucinate a lot of random things in the response. I think it needs to be shortened.
Updated the system prompt to actually use the system key in the POST request to the LLM. After also updating the prompt slightly, it seems to be hallucinating less.
So much pain. LLM keeps hallucinating. Just gets a lot of things wrong in general. Including response examples seems to help a little. I'll probably try using a completely different LLM.