关键词: 词元经济, 高质量数据集, 数据产业, 人工智能, 大语言模型, 范式变迁

Abstract: [Purpose/Significance] Token economy has reshaped the value accounting logic of large model industries, and the traditional scale-prioritized dataset construction mode has led to a prominent efficiency paradox. This paper aims to clarify the logic behind the transformation of objective in the construction of high-quality datasets under the token economy, and establish a new full-process dataset construction mode adaptable to token efficiency constraints. [Design/Methodology] This paper constructs an "Objective-Mode" analysis framework, sorts out differentiated data requirements throughout the full lifecycle of large models, analyzes the lock-in effect of original institutional complementarities in each link of dataset construction, and deduces full-link transformation solutions covering data collection, cleaning, annotation and evaluation. [Findings/Conclusion] The core objective of dataset construction has shifted from merely expanding data volume to maximizing the marginal value per token. Under new economic constraints, the original collaborative division system across all dataset development stages have fallen out of balance. Partial incremental enhancements, while beneficial, are insufficient to resolve structural contradictions. Consequently, a systematic transformation of the entire lifecycle is indispensable. Dataset construction needs to shift to a new operation system featuring targeted collection, semantic-level cleaning, pragmatic annotation, and dynamic token efficiency evaluation. [Originality/Value] This paper proposes a dynamic evaluation system centered on token consumption and inference benefits, and develops an integrated governance scheme covering the whole chain from data production to model inference. It can provide countermeasure references for government data standard formulation, enterprises datasets development, and industry data collaborative construction.

Keywords: Token economy, High-Quality datasets, Data industry, Artificial Intelligence, Large Language Models, Paradigm shift