AbijahKaj/qwen3-4b-skidl
AbijahKaj/qwen3-4b-skidl is a 4 billion parameter Qwen3-based language model fine-tuned by AbijahKaj to generate executable SKiDL Python netlists from natural language circuit descriptions. This model specializes in electronic circuit design, translating text prompts into functional Python code for KiCad, leveraging a 32768 token context length. It was trained on over 100,000 SKiDL Python circuit examples, achieving 95-96% token accuracy and an average functional score of 0.883 on held-out circuits. Its primary differentiator is its ability to directly produce KiCad-compatible netlists, streamlining the electronic design process.
Loading preview...
Model Overview
AbijahKaj/qwen3-4b-skidl is a specialized 4 billion parameter model, fine-tuned from the Qwen/Qwen3-4B base, designed for generating executable SKiDL Python netlists from natural language descriptions of electronic circuits. This model acts as an expert electronics engineer, translating prompts like "ESP32 board with BME280 on I2C and USB-C" into valid Python code using the SKiDL library, which can then be used to generate .net files for KiCad.
Key Capabilities & Differentiators
- Direct SKiDL Python Generation: Uniquely outputs functional Python code for circuit design, directly compatible with KiCad, bypassing the complexities of traditional netlist formats.
- Specialized for Electronic Design: Fine-tuned on a massive dataset of 100,179 SKiDL Python circuit examples, including conversions from LTspice and GitHub KiCad schematics.
- High Accuracy: Achieved 95-96% token accuracy and an evaluation loss of 0.1581 during training, demonstrating strong performance in generating correct SKiDL code.
- Efficient Circuit Representation: Leverages SKiDL's Python-based HDL, which is significantly more compact and LLM-friendly than KiCad's native s-expression netlists, leading to better generation performance.
Training Details
The model was fine-tuned using SFT + LoRA (r=64, \u03b1=32) over 2 epochs, targeting modules like q/k/v/o_proj and gate/up/down_proj. It utilized a maximum context length of 8192 tokens during training. Functional validation on held-out circuits showed an average score of 0.883, confirming its ability to produce correct and usable circuit designs.
Ideal Use Cases
This model is ideal for engineers, hobbyists, and developers looking to:
- Rapidly prototype electronic circuits from natural language descriptions.
- Automate the generation of KiCad-compatible netlists.
- Streamline the initial design phase of electronic projects by converting conceptual ideas into executable code.