SchGen

SchGen is a domain-specialized model and dataset framework for generating PCB schematics from natural-language descriptions. It introduces a semantic-grounded code representation for schematics and a dataset built from real-world open-source hardware designs, with an agentic sketch pipeline for dataset construction and a LoRA fine-tuning pipeline for GPT-oss models.

Generation runs from a dataset entry or from a direct user request, producing Python code in the schematic representation, which is then executed to create a KiCad project and schematic file. The dataset is built mainly from SparkFun resources under CC BY-SA 4.0 and includes KiCad schematics, semantic code representations, synthesized user requests, and chain-of-thought traces. The trained model is published on Hugging Face as microsoft/SchGen with the dataset as microsoft/SchGen_dataset.

The README is explicit that generated schematics may still contain incorrect connections, component selections, or layout inconsistencies, and that users should inspect the generated schematic in KiCad before proceeding to layout. The disclaimer states generated schematics and layouts should always be reviewed by experienced engineers before fabrication or deployment, and that SchGen does not guarantee electrical correctness or manufacturability.