distill-classifier

Teacher-labeler hard-label distillation reduces classifier cost by using multiple frontier models to generate consensus labels, which then train a lightweight open-weight student model. This approach maintains high accuracy for knowledge-bound classification tasks while significantly lowering inference latency and operational deployment costs.

Multi-teacher consensus labeling supports binary, multi-class, and multi-label classification tasks, alongside structured extraction. This workflow leverages multiple teacher models to curate high-quality datasets for supervised fine-tuning of a lightweight student model.

Yes, you can evaluate a distilled student classifier against its baseline teacher model using the built-in AB test framework. This comparison measures the student's performance and informs the deployment gating system to ensure accuracy is maintained before replacing the frontier classifier.

You should use the four-way verdict system—PROMOTE, SHADOW, COLLECT, or STOP—when deciding whether to deploy a distilled student classifier. It provides a data-driven gate based on AB test results to control deployment costs and ensure the lightweight model meets accuracy requirements.

Yes, building a teacher-student distillation workflow requires multiple frontier models to perform consensus labeling. These teacher models generate the majority voting labels needed to curate the supervised fine-tuning data for training the lightweight open-weight student classifier.

The best way to replace expensive frontier classifiers for structured extraction is through hard-label distillation. By training a lightweight open-weight student on consensus-labeled data from multiple teachers, you achieve comparable accuracy with reduced latency and lower inference cost.