According to the abstract of a new arXiv paper, data selection plays a central role in modern large language model training, as the quality of training data strongly influences downstream performance. The abstract also cites a common approach of curating a development set, though the excerpt ends before fully describing that approach.
The paper, titled ICLHerd, proposes a herding-based data selection method for LLMs, using in-context learning as a proxy. The title and abstract together indicate that the method uses in-context learning to guide the selection of training data, but the available excerpt does not include specifics on the algorithm or results.
Since only this single source was provided, there are no differing findings to compare. Full details of the method and its evaluation would require reading the complete arXiv paper.