LLM adaptation

SIEVE Wants Models to Learn From Three Examples, but the Trick Is Cutting Context

SIEVE uses SIEVE-GEN to create synthetic queries from decomposed context and then distills them into model weights.