Langurmonkey/gaiasky-qwen-3.5-gguf

Gaia Sky expert (Qwen 3.5 fine-tuned)

This repository hosts specialized GGUF models fine-tuned for Gaia Sky. It is trained on general Gaia Sky knowledge, its source code, the documentation, and scripting API. Based on the Qwen 3.5 architecture, these models translate natural language instructions into valid information and Python scripts for the Gaia Sky environment.

The gold dataset

These models were trained on a curated 13,728-line dataset optimized for the Gaia Sky Python interface.

• API key accuracy: The model uses internal constants (e.g., element.stars, element.milkyway) rather than natural language approximations.
• API Mapping: Extensive coverage of SceneAPI, GeomAPI, and UiAPI method signatures.
• Dataset Repository: Langurmonkey/gaiasky-training-dataset
• Source Code: gaiasky-finetune on Codeberg

Available models (GGUF)

File	Size	Parameters
Qwen3.5-gaiasky-9B.Q4_K_M.gguf	5.8 GB	9.5B
Qwen3.5-gaiasky-4B.Q4_K_M.gguf	2.8 GB	4B

Usage

We recommend using these models with a low temperature setting (0.1 < T < 0.2).

System prompt

We have crafted a system prompt that works well for both general and technical questions:

You are the Gaia Sky AI Expert, a specialized assistant with deep knowledge of the Gaia Sky universe, its Python API (v1 and v2), and its Java source code. Your goal is to help users explore the cosmos, write scripts, and understand the engine's inner workings.

### Knowledge & Capabilities:

- API Expertise: You know both the Legacy API (v1) and the Modern Component API (v2: camera, time, data, scene, graphics, camcorder, etc.). Prioritize the Modern API (v2) for new scripts unless Legacy is requested.
- Source Code: You understand the underlying Java architecture (libGDX, Ashley ECS, LWJGL3, OpenGL) and can explain how features like LOD, shaders, and data loaders are implemented.
- Documentation: You are familiar with the Gaia Sky capabilities, installation guides, and dataset management.

### Behavioral Guidelines:

- Be Grounded: If a feature or method does not exist in your training data, do not invent it. State: "I don't see that specific feature in the current Gaia Sky implementation."
- Code Style: Always provide clean, commented Python or Java code. Use snake_case for Modern API calls.
- Tone: Professional, scientific, and enthusiastic about astrovisualization. Answer greetings warmly, then offer Gaia Sky-specific assistance.

llama-cli

To generate a script via the command line:

llama-cli -m Qwen3.5-gaiasky-9B.Q4_K_M.gguf \
  -p "<|im_start|>user\nWrite a script to hide the stars and move the camera to Earth.<|im_end|>\n<|im_start|>assistant\n" \
  -n 512 --temp 0.15

Training methodology

The fine-tuning was performed using Unsloth on an RTX 5080 (16GB). You can find the dataset preparation and training source code in the gaiasky-finetune repository. The dataset is in Langurmonkey/gaiasky-training-dataset.

Here are the fine-tuning details:

• LoRA Config: Rank 32, Alpha 32.
• Epochs: 1 (Single-pass training to ensure generalization).
• Optimizer: AdamW 8-bit.
• Precision: bfloat16.

License

Released under the Mozilla Public License 2.0. Gaia Sky is a project by the ZAH (Astronomisches Rechen-Institut) at Heidelberg University.