Dataset Curation Pipeline

Magpie synthetic generation costs 30x less than human curation. Always run benchmark contamination checks (MMLU, HumanEval, GSM8K n-gram overlap) before training — contaminated data inflates eval scores without real improvement. A 10K curated set beats 100K noisy data.