README.md

---
datasets:
- Sweaterdog/Andy-4-preview-1
- Sweaterdog/Andy-4-preview-2
language:
- en
base_model:
- unsloth/DeepSeek-R1-Distill-Llama-8B-bnb-4bit
tags:
- gaming
- minecraft
---

THIS MODEL HAS BEEN DEPRECATED, PLEASE MOVE TO THE LATEST [BASE MODEL](https://huggingface.co/Sweaterdog/Andy-4-vl-base-LoRA)

# 🧠 Andy-4-base 🧠

Welcome to **Andy-4-base** – a revolutionary new model designed for playing Minecraft via the Mindcraft framework. 

This AI is crafted to push the boundaries of gameplay, reasoning, and multi-language summarization, making it one of the most versatile and powerful models available for Minecraft enthusiasts.

# Overview

Andy-4-base is an 8B parameter model tuned from Llama3.1 8B. 

It has been trained over three weeks on a single RTX 3090 using two carefully curated datasets. 

The model underwent 2 epochs on the first dataset with a higher learning rate and 4 epochs on the second dataset with a much lower learning rate, ensuring a balanced and robust learning process. 

This training regimen, coupled with advanced techniques like manual learning rate adjustments and dynamic dataset modifications, has allowed Andy-4-base to outperform its competitors, including GPT-4o-mini, Claude 3.5 Haiku, Mineslayer v2, and the former leader, Andy-3.6.

# Important notes

This is a base model, meaning it was generally tuned to play Mindcraft, and some important info about Mindcraft, it was **NOT** designed to be the final model

## What is the difference between this and Andy-4?

Andy-4 will be a fine tuned version of Andy-4-base, with more up to date information, as well as more information about how to respond to what the user wants, rather than playing by itself forever

## Other Notes

Andy-4-base was trained on a small subset of usernames for the model to understand names besides `Andy`, however this caused the model to only know those names, if you plan on using this model, refer to the `usernames` file

Andy-4-base is **NOT A FINISHED MODEL!!** Again, I must reiterate this model is used for preliminary testing as well as public feedback on general performance.

# How to install

On Huggingface, press the `Use this model` dropdown menu, and choose `Ollama`, then in the drop down menu, choose your quantization, following this GPU VRAM chart:

*All of these values assume a context window size of 8192 or less*
```
F16 = 20+ GB
Q8_0 = 12+ GB
Q5_K_M = >8 GB
Q4_K_M = 8+ GB
Q3_K_M = 6-8 GB on Minecraft LOW settings
Q2_K = 6-8 GB on Minecraft LOW settings
```
If you do not have a dedicated GPU, there is a [guide in the Mindcraft discord](https://ptb.discord.com/channels/1303399789995626667/1347027684768878644/1347027684768878644) server for setting up Cloud Computing for free

If you want a more custom install, go to the `files` tab on Huggingface, and download the quantization you want and the `Modelfile`

Once downloaded, open `Modelfile` in a text editor, and change the `FROM` tag to the exact path of the GGUF weights you installed without the quotes, such as `C:\users\jimmy\downloads\Andy-4-base.Q3_K_M.GGUF`

Open the directory of your Modelfile in the terminal, or command prompt, and then run `ollama create Andy-4-base -f Modelfile`, which will create `Andy-4-base` as the model, based on the Modelfile, which includes some important information such as system prompt, chat template, context length, as well as sampling settings.

For most people, using the direct HF method of downloading Andy-4-base is recommended, but you will suffer from a shorter context window, and possibly sparatic text.

For the full Andy-4 model, there will be a direct download on Ollama, so you won't have to do it this method.

## Installation notes

Andy-4-base, and Andy-4, will support a context window of up to `131072` tokens, but to run the Q5_K_M version with that context length, at stock settings on ollama, you would need an RTX 5090, and you would have 3GB of VRAM left over for Minecraft

A context window of `8192` is smaller, but it can still allow for great conversations with the model, and since the model knows all of the commands by heart, you can cut the command docs from the system prompt to lower context usage.

If you have a 6GB GPU, and want to run it locally, you will need to use the advanced installation method, as well as set the context length in the modelfile to `4096` to allow adequate VRAM for Minecraft.

If you want to know your specific VRAM usage, for Context length as well as the model, follow [This Huggingface Repo](https://huggingface.co/spaces/NyxKrage/LLM-Model-VRAM-Calculator)

If you want to have a larger context window, for less memory, and are okay with some hiccups when it comes to remembering things, follow this guide:

### On Windows:

1. Close Ollama
2. Open System properties by searching for `Edit the system environment variables` in the windows start menu
3. In the bottom left, click on `Environment Variables...`
4. Navigate to `System Variables` and press `New...`
5. Name the variable `OLLAMA_FLASH_ATTENTION` and set the value to `1`
6. Make *another* variable, name it `OLLAMA_KV_CACHE_TYPE` and set the value to `q8_0`, but if you are okay with more instability, and want more VRAM savings, set it to `q4_0`
7. Press `Okay` in the bottom, and then close out of the `System Properties` window
8. Now you can start Ollama again, and have a quantized context window

### On Linux / MaxOS

1. Open terminal
2. Ensure Ollama is not active, as in no language models can be ran
3. Run the following command:
```bash
export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE="q8_0"
ollama serve
```
-Or if you want more context, and are okay with instability-
```bash
export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE="q4_0"
ollama serve
```

# Key Features

- **Revolutionary Performance:** Specifically engineered for Minecraft gameplay, providing creative, strategic, and efficient in-game decision-making.
- **Advanced Reasoning Capabilities:** If you include `Reason in your responses with <think> and </think>` in your prompt, or something similar, Andy-4-base will provide detailed reasoning to enhance performance, at the cost of speed.
- **Multi-Language Summarization:** Capable of summarizing content in multiple languages, making it more efficient at remembering it's history.
- **Building and Creativity:** Not only can it play Minecraft, but it also excels in constructing complex structures and solving intricate in-game challenges.
- **Vast Knowledge Base:** Possesses extensive knowledge about Minecraft, including game mechanics, strategies, and creative builds.
- **Open Source:** Completely open source and available to the community for further development and experimentation.
- **Versatile Utility:** Excels in building, reasoning, summarizing, and strategic gameplay.

# Open Source and Licensing

Andy-4-base is 100% open source and is licensed under the [Apache 2.0 License](LICENSE). We believe in transparency and community collaboration, and all source code and training details are available for review and contribution.

## LoRA Weights

Access the LoRA weights for Andy-4-base here: [Andy-4-base LoRA Weights](https://huggingface.co/Sweaterdog/Andy-4-base-LoRA)

# Datasets

The model was trained on two distinct datasets:
- [Dataset 1: Andy-4-Preview-1](https://huggingface.co/datasets/Sweaterdog/Andy-4-preview-1) (trained for 2 epochs with a higher learning rate)
- [Dataset 2: Andy-4-Preview-2](https://huggingface.co/datasets/Sweaterdog/Andy-4-preview-2) (trained for 4 epochs with a much lower learning rate)

# Usage

To integrate Andy-4-base with the Mindcraft framework, adjust your configuration settings as needed and follow the execution instructions provided separately. 

Customize the personality in the `conversing` profile section and configuration files to optimize performance for specific in-game tasks or environments. 

Andy-4-base is designed to seamlessly integrate with your existing Mindcraft setups, enhancing both gameplay and creative capabilities.

# Mindcraft Project

Download and explore the Mindcraft project on GitHub:  
[Mindcraft Project on GitHub](https://github.com/kolbytn/mindcraft)

## How to use

On Mindcraft, ensure you put `ollama/` before the model name, this can be something like `ollama/hf.co/sweaterdog/andy-4-base:q5_k_m`

## Notes on Quantization

Typically, the smaller the quantization, such as Q2_K being the lowest, the more the model is to make a mistake, however the Q8_0 weights perform nearly identically to the F16 weights.

For most cases, I suggest Q5_K_M or Q4_K_M, this is for 8GB GPUs, anything less and I would follow the guide on how to use free cloud computing on the discord.

# Disclaimer

**Important Notice:**  
Andy-4-base is a preview model and, while it represents a significant advancement in AI-driven Minecraft gameplay, please be aware of the following:
- **Performance Variability:** Due to its experimental nature, the model may not always deliver ultimate performance in every scenario.
- **Ongoing Development:** This preview release is intended for early testing and community feedback. You might encounter occasional inconsistencies or limitations as further refinements are made.

# Community and Contact

Join our vibrant community for discussions, support, and feedback:
- **Discord (Mindcraft Server):** [Join the Mindcraft Discord](https://discord.gg/NGfCGs7aXJ)
- **Huggingface Page:** Explore more experimental models and projects: [Sweaterdog on Huggingface](https://huggingface.co/Sweaterdog)
- **Discord Username:** `Sweaterdog`

We welcome your feedback, suggestions, and contributions as we continue to improve Andy-4-base and push the boundaries of AI in gaming.

# Acknowledgements

We extend our gratitude to all the developers and researchers who have contributed to the evolution of AI in gaming.

```
## License
This model is licensed under the Andy 1.0 License.  
Credit: https://huggingface.co/Sweaterdog  
Acknowledgment: This work uses data and models created by @Sweaterdog.
```