| --- |
| base_model: Qwen/Qwen3-8B |
| library_name: peft |
| pipeline_tag: text-generation |
| tags: |
| - lora |
| - qlora |
| - peft |
| - transformers |
| - godot |
| - game-development |
| - code-generation |
| - text-generation |
| language: |
| - en |
| license: apache-2.0 |
| --- |
| |
| # SoloLabs GameForge 1 |
|
|
| **SoloLabs GameForge** is a QLoRA adapter for [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B), fine-tuned to assist with complete Godot 4 game creation from natural-language prompts. It is intended to generate and explain GDScript, Godot scene trees, project structure, 3D-game mechanics, asset manifests, and Web-preview-oriented build instructions. |
|
|
| This repository contains the **LoRA adapter**, not the full Qwen3-8B base-model weights. The base model is downloaded separately by Transformers from `Qwen/Qwen3-8B`. |
|
|
| ## Model details |
|
|
| | Field | Value | |
| |---|---| |
| | Model name | SoloLabs GameForge 1 | |
| | Fine-tuning method | QLoRA / PEFT LoRA | |
| | Base model | Qwen/Qwen3-8B | |
| | LoRA rank | 16 | |
| | LoRA alpha | 32 | |
| | LoRA dropout | 0.05 | |
| | Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj | |
| | Training records | 8,000 total; 6,400 train, 800 validation, 800 test | |
| | Intended domain | Godot 4 code and 3D game project generation | |
| |
| ## Intended use |
| |
| Use the adapter for prototyping Godot 4 projects, generating GDScript and scene-tree plans, drafting 3D-game mechanics, creating asset manifests, and explaining how to validate a generated project. The associated SoloLabs pipeline adds structural project checks, Godot Web export, headless Playwright preview verification, and allowlisted public asset discovery. |
| |
| The model should be treated as a coding assistant and project generator. Generated code must be reviewed, tested, and security-checked before use. Asset downloads must be independently reviewed for license compatibility before redistribution. |
| |
| ## Important limitations |
| |
| The adapter is not a standalone 8B model; it must be loaded on top of Qwen3-8B. The model's native output can use a design-object schema, while the production materializer expects a canonical file-list schema. The validated pipeline detects this mismatch, preserves the raw generation, and can use a deterministic procedural fallback for execution. |
| |
| The Kaggle validation demonstrated a working Godot Web export and headless Playwright verification. Android APK export remained blocked by a Godot 4.4.1 headless Android-preset validation problem in the Kaggle environment. This model card does not claim that every prompt produces a finished commercial game or a guaranteed APK. |
| |
| ## Loading the adapter |
| |
| ```python |
| import torch |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| from peft import PeftModel |
| |
| base_id = "Qwen/Qwen3-8B" |
| adapter_id = "teja123098/sololabs.1" |
| |
| tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True) |
| base = AutoModelForCausalLM.from_pretrained( |
| base_id, |
| torch_dtype=torch.float16, |
| device_map="auto", |
| trust_remote_code=True, |
| ) |
| model = PeftModel.from_pretrained(base, adapter_id) |
| model.eval() |
| |
| prompt = "Create a Godot 4 third-person 3D arena shooter with a player, two weapons, enemies, pickups, and a Web preview. Return a complete project file manifest with GDScript and scene trees." |
| inputs = tokenizer(prompt, return_tensors="pt").to(model.device) |
| with torch.no_grad(): |
| output = model.generate(**inputs, max_new_tokens=2048, do_sample=False) |
| print(tokenizer.decode(output[0], skip_special_tokens=True)) |
| ``` |
| |
| For low-memory inference, load the base model using a compatible 4-bit `BitsAndBytesConfig` and then attach this adapter with `PeftModel.from_pretrained`. |
|
|
| ## Training summary |
|
|
| The adapter was trained on Kaggle using QLoRA with 4-bit NF4 quantization, double quantization, FP16 compute, LoRA rank 16, alpha 32, dropout 0.05, and the seven attention/MLP projection target-module groups listed above. The dataset contains 8,000 structured examples covering Godot 4 project generation, GDScript, 3D gameplay, gun-game mechanics, scene trees, asset manifests, Web export, and validation workflows. |
|
|
| ## Responsible use |
|
|
| Do not use generated code to introduce malware, evade software protections, infringe third-party assets, or deploy unsafe systems. Review all generated scripts and dependencies. Treat web-discovered assets as untrusted until their source, license, checksum, and compatibility have been verified. |
|
|
| ## Related resources |
|
|
| - Base model: [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) |
| - Godot Engine: [godotengine.org](https://godotengine.org/) |
| - Playwright: [playwright.dev](https://playwright.dev/) |
| - SoloLabs GameForge Kaggle notebook: [tej8789/notebookcbedb4fe86](https://www.kaggle.com/code/tej8789/notebookcbedb4fe86) |
| - Training adapter source dataset: [tej8789/sololabs-gameforge-qwen3-8b-lora-v8](https://www.kaggle.com/datasets/tej8789/sololabs-gameforge-qwen3-8b-lora-v8) |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{sololabs_gameforge_2026, |
| title = {SoloLabs GameForge 1}, |
| author = {SoloLabs}, |
| year = {2026}, |
| note = {QLoRA adapter for Qwen3-8B specialized in Godot 4 game generation} |
| } |
| ``` |
|
|
