Instruction Backtranslation & Self-Alignment (Humpback / Kun)

Trains an inverse model $\pi_{\text{inv}}(x \mid y)$ to predict candidate user instructions $x$ from raw, unlabelled web/textbook documents $y$, then uses the forward model to score and filter $(x, y)$ pairs by quality, turning massive unlabelled corpora (like Wikipedia, StackExchange, or arXiv) into high-quality instruction-tuning datasets without distilling from a proprietary teacher.

Preserves human-written ground-truth diversity in responses $y$, avoiding the homogenizing "AI slop" tone of purely forward-generated synthetic responses.