Use Frontier Agents to Build Small Classifiers for Cheap, Large-Scale Data Curation
Key Info
A developer used frontier AI agents to build a small document-purpose classifier from just 200 agent-labeled examples, then applied it to curate 191,724 educational PDFs for roughly $0.70.
Highlights
- Combines frontier agents, SetFit, and Hugging Face Jobs to create a lightweight, reusable classifier.
- Demonstrates a “tokenmaxxing” workflow: use powerful models to generate labels, then deploy cheap small models for large-scale inference.
- Categorized 191,724 FinePDFs-Edu documents at a cost of about $0.70, showing large-scale data curation doesn't require expensive model calls.