🧠 Andy‑4 Family

A unified repository for the Andy‑4 family of specialist AI models, each tuned for enhanced Minecraft gameplay and multimodal capabilities* via the Mindcraft framework.

*Multimodal variants are coming, and are not yet released

---

Shared Metadata

• Language: English (en)
• Tags: gaming, minecraft, mindcraft
• License: Andy 1.1 License

---

🧠 Andy‑4 ⛏️

Andy‑4 is an 8 billion‑parameter specialist model tuned for Minecraft gameplay via the Mindcraft framework. Trained on a single NVIDIA RTX 3090 over three weeks, Andy‑4 delivers advanced reasoning, multi‑step planning, and robust in‑game decision‑making.

⚠️ Certification: Andy‑4 is not yet certified by the Mindcraft developers. Use in production at your own discretion.

🔍 Model Specifications

• Parameter Count: 8 B
• Training Hardware: 1 × NVIDIA RTX 3090
• Training Duration: ~3 weeks
• Data Volumes:
  • Messages: 179,384
  • Tokens: 425,535,198
  • Conversations: 62,149
• License: Andy 1.1 License
• Repository: Andy-4

Datasets

datasets:
  - Sweaterdog/Andy-4-base-1
  - Sweaterdog/Andy-4-base-2
  - Sweaterdog/Andy-4-ft
language:
  - en
base_model:
  - unsloth/Llama3.1-8B
tags:
  - gaming
  - minecraft
  - mindcraft

📊 Training Regimen

1. Andy‑4‑base‑1 dataset (47.4 k examples)

   • Epochs: 2
   • Learning Rate: 7 × 10⁻⁵

2. Andy‑4‑base‑2 dataset (48.9 k examples)

   • Epochs: 4
   • Learning Rate: 3 × 10⁻⁷

3. Fine‑tune (FT) dataset (4.12 k examples)

   • Epochs: 2.5
   • Learning Rate: 2 × 10⁻⁵

• Optimizer: AdamW_8bit with cosine decay
  
• Quantization: 4‑bit (bnb-4bit) for inference
  
• Warm‑Up Steps: 0.1% of each dataset

🚀 Installation

First, choose your quantization (context window base: 8192):

Quantization	VRAM Required
F16	16 GB+
Q5_K_M	8 GB+
Q4_K_M	6–8 GB
Q3_K_M	6 GB (low)
Q2_K	4–6 GB (ultra)

1. Installation via Ollama

1. Select your desired quantization
   
2. Copy the provided ollama run command
   
3. Execute it in your terminal
   
4. Use the model (e.g., ollama/sweaterdog/andy-4:latest)

2. Manual Download & Setup

1. Download:

   • From the Hugging Face Files tab, download the .GGUF weights (e.g., Andy-4.Q4_K_M.gguf).
     
   • Retrieve the Modelfile from this repo.

2. Configure Modelfile:

   FROM /path/to/Andy-4.Q4_K_M.gguf
   

   Optional: Adjust num_ctx for extended context windows if you have sufficient VRAM.

3. Register Locally:

   ollama create andy-4 -f Modelfile
   

If you lack a GPU, refer to the Mindcraft Discord guide for free cloud options.

🔧 Context‑Window Quantization

To reduce VRAM usage for context caching:

Windows

1. Close Ollama.
   
2. In System Properties → Environment Variables, add:

   
   OLLAMA_FLASH_ATTENTION=1  
   OLLAMA_KV_CACHE_TYPE=q8_0   # or q4_0 (more savings, less stable)
   

3. Restart Ollama.

Linux/macOS

export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE="q8_0"   # or "q4_0"
ollama serve

📌 Acknowledgments

• Data & Models By: @Sweaterdog
  
• Framework: Mindcraft
  
• LoRA Weights: https://huggingface.co/Sweaterdog/Andy-4-LoRA

---

🤏 Andy‑4‑micro 🧠

Andy‑4‑micro is a lightweight, 1.5 B‑parameter variant of Andy‑4, optimized for responsive local inference and experimentation within the Mindcraft framework.

💡 Trained on a single NVIDIA RTX 3070 over four days, Andy‑4‑micro maintains strong performance while staying efficient.
⚠️ Certification: Not yet certified by Mindcraft developers. Use at your own discretion.

📊 Model Overview

• Parameter Count: 1.5 B
  
• Training Hardware: 1 × NVIDIA RTX 3070
  
• Training Duration: ~4 days
  
• Total Tokens: ~42 M
  
• Base Architecture: Qwen 2.5
• License: Andy 1.1 License
  
• Repository: Andy-4-micro

Datasets

datasets:
  - Sweaterdog/Andy-4-base-2
  - Sweaterdog/Andy-4-ft
language:
  - en
base_model:
  - unsloth/Qwen2.5-1.5B-Instruct-unsloth-bnb-4bit
tags:
  - gaming
  - minecraft
  - mindcraft

🚀 Installation

First, choose your quantization (context window base: 8192):

Quantization	VRAM Required
F16	5 GB
Q8_0	3 GB+
Q5_K_M	2 GB+
Q3_K_M	1GB or CPU

1. Installation via Ollama

1. Select quantization
   
2. Copy and run the ollama run command
   
3. Use ollama/sweaterdog/andy-4:micro-q8_0 locally

2. Manual Download & Setup

1. Download: Grab the .GGUF weights (e.g., Andy-4-micro.Q4_K_M.gguf) and Modelfile.
   
2. Configure Modelfile:

   
   FROM /path/to/Andy-4-micro.Q4_K_M.gguf
   

   Optional: Tweak num_ctx as needed.
3. Register:

   
   ollama create andy-4-micro -f Modelfile
   

Free GPU options: see the Mindcraft Discord guide.

🔧 Context‑Window Quantization

Use the same environment-variable tweaks listed above for AdamW-flash attention and KV cache.

📌 Acknowledgments

• Data & Model By: @Sweaterdog
  
• Framework: Mindcraft
  
• LoRA Weights: https://huggingface.co/Sweaterdog/Andy-4-micro-LoRA