Evol-Instruct
A method that expands instruction datasets by repeatedly rewriting seed instructions with a large language model. It carries a structural risk of inadvertently introducing benchmark test items into training data.
Timeline (last 90 days)
The Evol-Instruct method expands the dataset by repeatedly rewriting seed problems; if benchmark problems are mixed into the seed pool, all variants inherit their core problem structure, and exact-match filters cannot detect these rewritten contaminations.
典型的合成数据技术包括 Self-Instruct、Evol-Instruct 以及基于种子数据的扩写和改写
All Facts (2)
The Evol-Instruct method expands the dataset by repeatedly rewriting seed problems; if benchmark problems are mixed into the seed pool, all variants inherit their core problem structure, and exact-match filters cannot detect these rewritten contaminations.
50%Unverified典型的合成数据技术包括 Self-Instruct、Evol-Instruct 以及基于种子数据的扩写和改写
50%