A self-synthesis method for alignment data that prompts an aligned LLM "with nothing" — feeding only the chat template's pre-query tokens — to make the model emit both instruction and response (ICLR 2025). The resulting datasets are widely reused for SFT (magpie-align/magpie, ~873★). UW-led (Zhangchen Xu et al.).

Paper

Dataset

datasetpost-trainingnlpdata