Instruction-Tuning Data Selection
Overview
The field asks which examples, conversations, or task mixtures best teach a model to follow instructions, reason, and handle specialized tasks. Methods range from quality and difficulty filters to embedding-based coverage, gradient or influence scores, model-activation signals, and adaptive selection guided by validation results. Researchers apply these approaches to text, code, dialogue, and multimodal instruction tuning, often testing whether small selected subsets can match or exceed full-data training on task benchmarks. The central tension is that selection depends on the model, task, and training budget, while costly scoring and search—and possible shifts in safety or other behaviors—make reliable evaluation difficult.
Latest papers
45 papers9 papers in the last four weeks, up 29% on the four weeks before. 0.1% of all new papers.