Researchers Build Dataset to Train AI Models on Analog Circuit Design

Researchers have developed a comprehensive dataset and training methodology to teach large language models analog circuit design principles, a specialized domain that has historically been challenging for general-purpose AI systems.

Building the Training Dataset#

The team constructed their dataset by extracting content from high-quality analog circuit textbooks and breaking it into discrete learning units. They then used a multi-agent framework to generate structured question-thinking-solution-answer quadruples—a format that captures both the final solution and the reasoning process behind it. The resulting dataset comprises 7.26 million tokens of unlabeled text for general pre-training and 112.65 million tokens of labeled examples for supervised fine-tuning on specific tasks.

Training Approach and Optimization#

Rather than starting from scratch, the researchers found that instruct-tuned models served as the best starting point for this specialized task. Their experiments revealed that while pre-training on general analog circuit text provided some benefit, the core gains came from fine-tuning on labeled examples. A key finding: combining supervised fine-tuning with KL divergence regularization—a technique that prevents models from drifting too far from their original training—yielded a 2.71 percentage-point improvement over standard fine-tuning alone. The team also documented practical implementation methods for training in resource-constrained settings.

Performance and Real-World Application#

A 32-billion-parameter model trained with this dataset achieved 84.59% accuracy on the AMSBench-TQA benchmark, a 15.67 percentage-point jump from its baseline performance. Beyond multiple-choice assessments, the trained model demonstrated capability in designing operational amplifiers—a practical engineering task evaluated using the Atelier framework, suggesting the approach produces knowledge that translates to authentic circuit design work.