---
datasets:
- Sweaterdog/Andy-4-base
- Sweaterdog/Andy-4-ft
language:
- en
base_model:
- unsloth/Qwen3-8B-bnb-4bit
tags:
- gaming
- minecraft
- mindcraft
---

# 🧠 Andy‑4 ⛏️


![image/png](https://cdn-uploads.huggingface.co/production/uploads/66960602f0ffd8e3a381106a/raWYEDo2An1biTLXd5PfN.png)
**Andy‑4** is an 8 billion‑parameter specialist model tuned for Minecraft gameplay via the Mindcraft framework.  Trained on a single 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

- **Parameters:** 8 B  
- **Training Hardware:** 1 × NVIDIA RTX 3090  
- **Duration:** ~3 weeks total  
- **Data Volumes:**  
  - **Messages:** 179,384  
  - **Tokens:** 425,535,198  
  - **Conversations:** 62,149  

- **Base Architecture:** Qwen 3 8b  
- **License:** [Andy 1.1 License](LICENSE)  
- **Repository:** https://huggingface.co/Sweaterdog/Andy‑4

---

## 📊 Training Regimen

1. **Andy‑4‑base‑1** dataset  
   - **Epochs:** 3  
   - **Learning Rate:**   7e-5
   - **Dataset Size:** 47.4k

3. **Fine‑tune (FT) dataset**  
   - **Epochs:** 2.5  
   - **Learning Rate:** 2e-5
   - **Dataset Size:** 4.12k

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

---

## 🚀 Installation

First, you need to choose your quantization, this chart is with the base of `8192` set as the context window

| 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 directly on Ollama

1. Visit [Andy-4 on Ollama](https://ollama.com/Sweaterdog/Andy-4)
2. Copy the command after choosing model type / quantization
3. Run the command in the terminal
4. Set the profile's model to be what you installed, such as `ollama/sweaterdog/andy-4:latest`

### 2. Manual Download & Modelfile

1. **Download**  
   - From the HF **Files** tab, grab your chosen `.GGUF` quant weights (e.g. `Andy-4.Q4_K_M.gguf`).  
   - Download the provided `Modelfile`.


2. **Edit**

   Change
   ```text
   FROM YOUR/PATH/HERE
   ```
   to
   ```text
   FROM /path/to/Andy-4.Q4_K_M.gguf
   ```
  *Optional*:
  Increase the parameter `num_ctx` to a higher value for longer conversations if you:

  **A.** Have extra VRAM

  **B.** Quantized the context window

  **C.** Can use a smaller model

3. **Create**  
   ```bash
   ollama create andy-4 -f Modelfile
   ```

This registers the **Andy‑4** model locally.

---

If you lack a GPU, check the [Mindcraft Discord guide](https://ptb.discord.com/channels/1303399789995626667/1347027684768878644/1347027684768878644) for free cloud setups.


## 🔧 Context‑Window Quantization

To lower VRAM use for context windows:

#### **Windows**

1. Close Ollama.  
2. In **System Properties → Environment Variables**, add:  
   ```text
   OLLAMA_FLASH_ATTENTION=1  
   OLLAMA_KV_CACHE_TYPE=q8_0   # or q4_0 for extra savings, but far more unstable
   ```  
3. Restart Ollama.

#### **Linux/macOS**

```bash
export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE="q8_0"   # or "q4_0", but far more unstable
ollama serve
```

---

## 📌 Acknowledgments

<details>
<summary>Click to expand</summary>

- **Data & Models by:** @Sweaterdog  
- **Framework:** Mindcraft (https://github.com/kolbytn/mindcraft)  
- **LoRA Weights:** https://huggingface.co/Sweaterdog/Andy-4-LoRA  
</details>

---

## ⚖️ License

See [Andy 1.1 License](LICENSE).  

*This work uses data and models created by @Sweaterdog.*