README.md

---
license: apache-2.0
tags:
  - geometry-dash
  - level-generation
  - gguf
  - tinyllama
base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
pipeline_tag: text-generation
---

# EditorAI

The main EditorAI model — a fine-tuned TinyLlama-1.1B-Chat that generates Geometry Dash levels as JSON with blocks, spikes, platforms, triggers, groups, color channels, and more.

Part of the [EditorAI](https://github.com/Entity12208/EditorAI) project — an AI-powered level generator mod for Geometry Dash.

## About EditorAI

[EditorAI](https://github.com/Entity12208/EditorAI) is a [Geode](https://geode-sdk.org) mod for Geometry Dash that lets you describe a level in plain text and have AI build it in the editor. It supports 8 AI providers (Gemini, Claude, OpenAI, Mistral, HuggingFace, Ollama, LM Studio, llama.cpp) and features blueprint preview, feedback learning, 15+ trigger types, and an in-game settings UI.

## Model Details

- **Base model:** TinyLlama-1.1B-Chat-v1.0
- **Training:** QLoRA (4-bit, rank 8) on 2368 examples (368 expert-crafted + 2000 synthetic), 2 epochs
- **Features:** Blocks, spikes, platforms, color triggers, move triggers, alpha triggers, rotate triggers, toggle triggers, pulse triggers, spawn triggers, stop triggers, speed portals, groups, color channels
- **GGUF quantization:** q4_k_m (637 MB)
- **Tested:** 8/8 tests passed, generates all trigger types

## Files

| File | Size | Description |
|------|------|-------------|
| `model.safetensors` | 2.1 GB | Merged fp16 model weights |
| `editorai-latest.gguf` | 637 MB | Quantized GGUF (q4_k_m) for llama.cpp / LM Studio |
| `config.json` | — | Model architecture config |
| `tokenizer.json` | — | Tokenizer |

## Setup

This model uses the **Zephyr/ChatML** chat template and works best with the following system prompt:

```
You are a Geometry Dash level designer. Return ONLY valid JSON with an analysis string and objects array. Each object needs type, x, y. Y >= 0. X uses 10 units per grid cell.
```

> **Recommended:** Use the Ollama version (`entity12208/editorai:latest`) which has the system prompt and template pre-configured.

## Usage with llama.cpp

```bash
wget https://huggingface.co/EditorAI-Geode/EditorAI/resolve/main/editorai-latest.gguf

llama-server -m editorai-latest.gguf --port 8080 --chat-template chatml

# In the EditorAI mod: set provider to "llama-cpp", URL to http://localhost:8080
```

## Usage with LM Studio

1. Download `editorai-latest.gguf` from this repo
2. Load it in LM Studio, set **Prompt Format** to **ChatML**
3. Set the **System Prompt** to the prompt above
4. Start the server
5. In the EditorAI mod: set provider to "lm-studio", URL to `http://localhost:1234`

## Usage with Ollama (recommended)

```bash
ollama pull entity12208/editorai:latest
```

In the EditorAI mod: set provider to "ollama" and select `entity12208/editorai:latest`.

## Output Format

```json
{
  "analysis": "A medium modern level with color transitions and moving platforms.",
  "objects": [
    {"type": "block_black_gradient_square", "x": 0, "y": 0, "color_channel": 10},
    {"type": "spike_black_gradient_spike", "x": 50, "y": 0},
    {"type": "color_trigger", "x": 80, "y": 0, "color_channel": 1, "color": "#0066FF", "duration": 1.5},
    {"type": "move_trigger", "x": 90, "y": 0, "target_group": 1, "move_x": 0, "move_y": 20, "duration": 1.0, "easing": 1},
    {"type": "end_trigger", "x": 400, "y": 0}
  ]
}
```

## Model Comparison

| Model | Size | Triggers | Quality | Speed |
|-------|------|----------|---------|-------|
| **editorai:latest** (this) | 637 MB | All types | Best | ~15-30s |
| editorai:mini | 379 MB | Color, move | Good | ~10-20s |

## Links

- **Mod:** [github.com/Entity12208/EditorAI](https://github.com/Entity12208/EditorAI)
- **Ollama:** [ollama.com/entity12208/editorai](https://ollama.com/entity12208/editorai)
- **Mini model:** [huggingface.co/EditorAI-Geode/editorai-mini](https://huggingface.co/EditorAI-Geode/editorai-mini)
- **Discord:** [discord.gg/5hwCqMUYNj](https://discord.gg/5hwCqMUYNj)

## License

Apache 2.0