v2: Human-verified training data — 5588 examples (1099 gold, 4378 reject, 105 silver, 29 calibration)

v2: Human-verified training data — 5583 examples (1099 gold, 4377 reject, 105 silver, 24 calibration)

MEGA curation dataset: 62,741 examples (2,450 vision, 60,291 text) — gold standard + curation queue + 71K gen-learning + prompts + knowledge

initial commit