MindCraft-Fine-Tuning 🧠🔨

Files needed for contributing to Fine tuned models.

Big models can play well, but can smaller models beat them?

We are setting out to make the best, and most affordable models for YOUR MindCraft experience

How can I help?

If you would like to help collect data, download the logger.js file, and whatever API endpoint you plan on using, such as claude.js DO NOT USE FINE TUNED LOCAL MODELS FOR MINDCRAFT FOR COLLECTING DATA

What do I do once I have installed logger.js, and my API endpoint?

After you downloaded logger.js, as well as all of the endpoints you would like to use, choose what model you will be using, if you want to add reasoning logs, ONLY Deepseek-R1 and it's distilled models, like LLaMa3.3 Deepseek-R1 distill or the Qwen2.5 32B distill, if you want vision, choose any multimodal model capable of vision inputs, if you want regular, choose any model besides deepseek-r1 and it's distills.

Once you have changed logger.js to the logging settings, put logger.js into the root file of MindCraft. Once logger.js is in the root file, navigate to src\models and DELETE the old API endpoint, such as gpt.js and insert the new API endpoint, such as hyperbolic.js

I finished those changes, what now?

After you have made the changes, there is nothing you need to do! Simply load up Minecraft, play with your bot, and the data will get collected automatically! Once you play for long enough, ~10 hours of playing or more, make a new dataset repository, on huggingface, and post a message to this channel with a link to your dataset repository. MAKE SURE TO CHANGE THE NAME OF THE FILE TO YOUR USERNAME BEFORE MAKING THE PULL REQUEST, OR IT WILL BE DENIED. Example: Beethoven_reasoning.csv or Duolingo_vision.csv or Jimmy_regular.csv

I have uploaded my logs to the folder, what now?

Now that you have uploaded your logs, there is nothing else you need to do! You can continue to collect more data, or not!

Official TODO list:

TODO List for LLM Fine-tuning

Make a larger dataset

Of course, more examples, the better the model, will ALWAYS be a work in progress

Develop a dataset including vision

A Far off need, Andy-vision will eventually be made

Make a reasoning dataset

Uses Deepseek-R1, or R1 distills to generate thought, to provoke thinking among Andy models, which may improve Andy-mini performance

Find GPU's here to tune the models

Tuning any model, is very computationally expensive, it took 3 hours to train Andy-3.5 on one epoch using an A100, having someone here provide the resources to do it, both for free and without a time limit can allow us to make the models better

Make Andy-Ultra

A model which can play Minecraft flawlessly, can reason and has vision