Alignment Data — Curation + Sources

Why it matters

Alignment data quality determines model quality. Curation shapes results.

The architecture

Sources: human, LLM-judge, synthetic.

Curation: filter noise + bias.

How it works end to end

Datasets: UltraFeedback, HH-RLHF, PKU-SafeRLHF.

Curation: dedupe, safety filter, bias check.

Balance: helpfulness + harmlessness.