Cheap Synthetic Math Data Boosts a Tiny 0.6B Model's Grade-School Math Accuracy From 36% to 49%

Fine-tuning small local models on 21,250 GPT-5-mini-generated math word problems, done entirely on a MacBook, jumped GSM8K accuracy from 36.5% to 49.1% for a 0.6B model and 53.5% to 66.5% for a 1.7B…