README.md

---
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}
}
```