Self-guided data selection is an automated machine learning technique in which a model evaluates and curates its own training or fine-tuning samples based on its internal capabilities, eliminating the need for manual human filtering or external evaluator models. In the context of instruction tuning for large language models, this approach assesses candidate data by quantifying the disparity between expected responses and the model's intrinsic generation behavior, thereby isolating the most informative and challenging examples. By autonomously filtering extensive datasets into a compact subset of high-value samples, self-guided data selection reduces computational and curation costs while preserving or improving overall training performance.
