Kodr — the game generator AI

Kodr is an open-source AI that turns design ideas into working games: engine code, gameplay systems, and complete levels — for any engine (Roblox Luau, Unity C#, Godot GDScript, Unreal C++).

It is trained from one of the Qwen3-class coder reference models using QLoRA, then merged into a single standalone checkpoint. The reference model is removed completely after training — only Kodr remains.

What it does

Capability	Example output
Code	Full scripts: controllers, combat, inventories, saving, round systems, AI
Maps	kodr-map-v1 JSON — heightmap, spawns, lighting, rules, objects
Any engine	Roblox Luau, Unity C#, Godot GDScript, Unreal C++
Compile maps	python model/tools/compile_map.py map.json --engine roblox

Quickstart

pip install -r model/requirements.txt

# chat
python - <<'PY'
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
m = AutoModelForCausalLM.from_pretrained("Radinkazemian/kodr", device_map="auto", torch_dtype=torch.float16)
t = AutoTokenizer.from_pretrained("Radinkazemian/kodr")
msg = [{"role":"system","content":"You are Kodr, a game generator AI."},
       {"role":"user","content":"A parkour arena with wall-jumps and a lava ring."}]
inp = t.apply_chat_template(msg, tokenize=True, add_generation_prompt=True, return_tensors="pt")
print(t.decode(m.generate(inp, max_new_tokens=800)[0].tolist(), skip_special_tokens=True))
PY

Train it yourself: run model/train.py (or the Colab notebook in this repo).

Architecture

• Reference: a Qwen3-class 27B coder model (open weights, Apache-2.0)
• Training: QLoRA (4-bit) + LoRA merge → one standalone .safetensors model
• Total params: 27B | Seq len: 1024

License

Apache-2.0. The dataset is small and hand-curated per engine; see model/dataset/build_dataset.py.