atsmt/rect2sheet-qwen-7b
atsmt/rect2sheet-qwen-7b is a 7.6 billion parameter Qwen2.5-Coder-7B-Instruct model fine-tuned by atsmt for generating sheet-metal solutions. It specializes in producing fold sequences, bends, and tab geometry from JSON descriptions of connected rectangular tabs. The model, with a 32,768 token context length, is optimized for data-driven sheet-metal design generation and research into constrained CAD generation.
Loading preview...
Rect2Sheet Qwen 7B Overview
atsmt/rect2sheet-qwen-7b is a specialized 7.6 billion parameter language model, fine-tuned from unsloth/qwen2.5-coder-7b-instruct. Its primary function is to generate sheet-metal solutions, including fold sequences, bends, bend directions, and resulting tab geometry, based on JSON descriptions of rectangular tab layouts. The model leverages a Qwen2ForCausalLM architecture and supports a substantial context length of 32,768 tokens.
Key Capabilities
- Sheet-Metal Solution Generation: Translates JSON descriptions of connected rectangular tabs into detailed sheet-metal solutions.
- Specialized Training: Trained on 19,231 manually verified designs from the Rect2Sheet dataset, ensuring domain-specific accuracy.
- JSON Output: Designed to produce parsable JSON output for direct integration into workflows.
- Llama.cpp Support: Quantized GGUF files are provided for efficient local execution using
llama-cli.
Intended Use Cases
- Research in Data-Driven Design: Ideal for exploring data-driven approaches to sheet-metal design generation.
- Candidate Solution Generation: Useful for producing initial Rect2Sheet solutions from inputs conforming to the dataset schema.
- Baseline for CAD/Geometry Research: Can serve as a baseline model for research in constrained CAD generation and geometry generation.
Limitations
It's important to note that generated JSON may be malformed or not conform to the Rect2Sheet schema. Syntactically valid output does not guarantee geometric validity or manufacturability, and the model does not replace engineering checks for collision, unfolding-overlap, or thin segments. Performance outside the training data's represented tab counts, geometry, mount types, and topology is unknown, and no benchmark results are currently reported.