Atlas · GenAI 2026

Training Data Curation

Dataset curation for SFT/PEFT

conceptPeak: 2024Dataset CurationAI consensus: 2/3

Prerequisites

  • SFT dataset curation is data curation applied to training data — curation skills are the foundation

Recommended reference

Zhou et al. (2023) 'LIMA: Less Is More for Alignment' — shows data quality > quantity for SFT; seminal paper with 1000 examples

Notes from AI deep research

Anthropic Opus

LIMA paper: 1000 dobrych przykladow > milion slabych

OpenAI Deep Research

Jakość danych → wpływ na dostrajanie [OA#42]

Google Deep Think

Formatowanie jakościowych danych [G#61 partial]

Related skills