SchGen

20B-parameter LLM for PCB schematic generation from natural-language requests. Built on GPT-OSS-20B and fine-tuned on ~8K paired user-request / schematic-code samples synthesized via GPT drafting + human correction. Takes a circuit description as input (e.g., "a 1.8V regulated supply from VIN using an AP2112K LDO") and outputs Python code that renders into KiCad schematics.

Evaluated on Valid Circuits (executability), Spatial Violation (symbol/wire overlap), and Netlist Accuracy (connectivity vs ground truth) — outperforms frontier LLM baselines on schematic generation. 13,312-token context. MIT license. Companion paper Luo, Ma, Zhang, Qiu (arXiv 2605.30345).