vishnuOI/unity-coder-30b

Unity Coder 30B

A QLoRA fine-tuned adapter for Qwen3-Coder-30B-A3B-Instruct specialized for Unity game development in C#.

Training

• Base model: Qwen/Qwen3-Coder-30B-A3B-Instruct (30B MoE, 3B active params)
• Method: QLoRA (4-bit NF4, r=16, alpha=32, target: q/k/v/o/gate/up/down projections)
• Dataset: vishnuOI/unity-dev-instructions
• Training pairs: 16,604 Unity instruction pairs
• Sources: Unity docs (32K pairs scraped), Stack Overflow [unity3d], GitHub Unity C# repos
• Hardware: 2x NVIDIA A100 80GB PCIe

Capabilities

• Unity C# scripting (MonoBehaviour, ScriptableObjects, coroutines, events)
• XR/VR development (OpenXR, XR Interaction Toolkit, spatial anchors)
• Physics, animation, UI Toolkit, NavMesh
• URP/HDRP shaders and rendering
• DOTS/ECS/Burst/Jobs performance patterns
• Editor scripting and tooling

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen3-Coder-30B-A3B-Instruct"
adapter_id    = "vishnuOI/unity-coder-30b"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

messages = [
    {"role": "system", "content": "You are an expert Unity game developer."},
    {"role": "user",   "content": "Write a MonoBehaviour that spawns enemies at random positions."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=512, temperature=0.1, do_sample=True)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))