Notes
MIPROv2, GEPA and few-shot on one DSPy extractor
Four ways of preparing one extractor, 100 held-out rows, 3 seeds, with compile cost and cost per 100 inferences; then GEPA from 100 to 1,000 training rows.
By Choiril Kurniawan
· 9 min read
Traces, notes and logs from the runs.
Choiril Kurniawan
I build and run machine learning and data systems. These are traces, notes and logs from the runs and the experiments: what I measured, what changed, and what it cost.
About me