---
license: apache-2.0
base_model: Qwen/Qwen2.5-7B-Instruct
library_name: transformers
pipeline_tag: text-generation
tags:
- geometry-dash
- gguf
- llama-cpp
- ollama
- tool-use
- geode
- qwen2.5
language:
- en
---

# EditorAI v3 — 7B GD Level Designer

Major upgrade over v2. Fine-tune of **Qwen/Qwen2.5-7B-Instruct** for the [EditorAI Geode mod](https://github.com/Entity12208/EditorAI).

| | v2 (1.5B) | **v3 (7B)** |
|---|---|---|
| Base model | Qwen2.5-1.5B-Instruct | **Qwen2.5-7B-Instruct** |
| Native context | 8 K | **32 K** |
| Q4_K_M size | 941 MB | **4.46 GB** |
| Schema reliability | drifts under verbose runtime prompts | matches the mod's runtime distribution |
| Tool use | works w/ mod fallback | native, no fallback needed |
| Output format | JSON | EAS (preferred) + JSON fallback |

## Files

- `editorai-v3-Q4_K_M.gguf` (4.46 GB) — primary ship target
- `Modelfile.v3` — Ollama Modelfile, 32K ctx, Qwen2.5 tool template

## Quick start

```bash
ollama pull entity12208/editorai:v3-7b
ollama create entity12208/editorai:v3-7b -f Modelfile.v3   # alternative
./llama-server -m editorai-v3-Q4_K_M.gguf -c 32768 --jinja  # llama.cpp
```

## Speed

- RTX 3050 6 GB (Ampere): ~25–35 t/s
- RTX 3090/4090: ~70–120 t/s
- Apple M3 Pro (Metal): ~30–45 t/s

## Training

- QLoRA 4-bit NF4, rank 16, alpha 32, lr 2e-4 cosine, paged_adamw_8bit
- A100 40 GB (Lightning.ai), ~58 min training (115 steps end of epoch 1)
- 3,700-row mixed dataset: 2,000 level-gen from real .gmd files, 1,200 multi-turn tool-use, 500 Alpaca instruct
- System prompt at training time matches the mod's runtime system prompt verbatim

## License

Apache-2.0, inherited from the base model.