Godoter-27B

What it's good at

• Writing idiomatic Godot 4 GDScript (nodes, signals, resources, @export/@onready, await, typed code).
• Building complete, multi-file systems: inventory, finite-state-machine AI, save/load, dialogue, multiplayer, shaders, combat (hitbox/hurtbox), navigation, crafting, skill trees, and more.
• Not confusing Godot 3 and Godot 4 APIs (move_and_slide(), instantiate(), CharacterBody2D/3D, source_color, FileAccess, signal .connect()/.emit(), @rpc, …).
• Staying in Godot instead of drifting to Unity, Python, web, or other frameworks.

Benchmark

Godoter vs. the base Qwen3.6-27B, both in the same Q6_K quant, same GPU, same prompts, thinking enabled for both. A task "passes" only if the answer uses the correct Godot 4 API and avoids the deprecated Godot 3 API (the check looks at code blocks, not prose).

On the hard set, reading the actual answers shows why the base struggles:

1. Framework drift — asked to "build a skill tree" or a "deck-builder", the base often replies in HTML/CSS/JS or Python; an RPC question gets answered in gRPC/Go; a lobby in WebRTC. It doesn't realize the question is about Godot. Godoter never drifts.
2. Lost in reasoning — on some prompts the base spends its whole token budget thinking and never reaches the code.

(The 68% even overstates the base, since the keyword rubric credits wrong-language or incomplete answers. The real gap is larger.)

How it was trained

• Base model: unsloth/Qwen3.6-27B
• Method: QLoRA, rank r = 16, alpha = 16, adamw_8bit, 4-bit base — adapter then merged to 16-bit.
• Dataset (~18,000 examples), all anchored to the official Godot 4 docs:
  • exact API reference extraction (methods, properties, signals with signatures)
  • documentation-grounded Q&A over the Godot 4 tutorials
  • 666 complete, multi-file Godot 4 systems authored as build / explain / extend triples
  • real Godot 4 GDScript completion pairs

Usage

This is a qwen3_5 / Qwen3.6 architecture model — it needs transformers 5.x.

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("Ruler97/Godoter-27B")
model = AutoModelForCausalLM.from_pretrained(
    "Ruler97/Godoter-27B", torch_dtype=torch.bfloat16, device_map="auto")

msgs = [{"role": "user", "content": "In Godot 4, make a CharacterBody2D move with the arrow keys."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=512)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

For local use, prefer the GGUF builds (LM Studio / llama.cpp / Ollama): Ruler97/Godoter-27B-GGUF (Q4 / Q6 / Q8).

Limitations

• It's a derivative of Qwen3.6-27B — ~99% of its capability is the base model; the fine-tune adds Godot-4 reliability, not new general reasoning.
• Knowledge is bounded by the training data and the base's cutoff; very recent Godot 4.x changes may not be reflected.
• The benchmark measures API correctness, not full architectural quality of generated code.
• Like any LLM, it can be confidently wrong — review generated code before shipping.

License & attribution

Godoter is a fine-tune of Qwen3.6-27B, which is licensed Apache 2.0, and Godoter is released under Apache 2.0 as well. Please keep attribution to the base model. Built with Unsloth.