qwen3-vl-cadquery

qwen3-vl-cadquery (fine-tuned)

This repository contains a fine-tuned Qwen3-VL model adapted for CadQuery-style CAD/code understanding and generation, trained on the ThomasTheMaker/cadquery dataset.

Model details

• Developed by: IFMedTechdemo
• Base model: unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit
• License: Apache-2.0
• Dataset: ThomasTheMaker/cadquery (~2000 examples)
• Training: 30 epochs
• Training stack: Unsloth + Hugging Face TRL

Intended use

This model is designed to help with CAD / parametric modeling workflows that resemble the CadQuery dataset distribution, including:

• Generating CadQuery-style Python code from natural-language CAD instructions.
• Explaining or modifying existing CadQuery snippets (e.g., “add 4 holes”, “increase wall thickness”, “fillet edges”, “change dimensions”).
• Basic reasoning about geometric intent described in text (dimensions, features, operations).

How to use (example)

Typical usage is as an instruction-following model where you provide:

• A natural-language task description, and optionally
• A partial CadQuery snippet to modify

Example prompt:

• “Create a parametric bracket with two mounting holes (M4), 3mm thickness, and a 20mm standoff. Output CadQuery code.”

Notes on training

• Fine-tuned for 30 epochs on ~2000 items from ThomasTheMaker/cadquery.
• Unsloth was used to speed up training and improve memory efficiency; TRL was used for the fine-tuning loop.

Deployment

This repo is tagged for text-generation-inference and transformers. You can run it with standard Transformers inference or deploy via TGI depending on your environment and quantization support.

Limitations

• The model can produce code that is not directly executable (syntax/import issues, missing variables, or invalid CadQuery API usage).
• Dimensions/constraints may be inconsistent; validate outputs before using them in production CAD pipelines.
• Complex assemblies, tight tolerance constraints, and manufacturing-ready outputs may require human review and iterative refinement.

License

Apache-2.0 (see metadata above).

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>