This article discusses the development of a specialized digital dataset for training intelligent syntactic analysis models for the Uzbek language. One of the main challenges in Uzbek Natural Language Processing (NLP) is the lack of high-quality, linguistically annotated corpora that account for the language's agglutinative nature. To address this, a 16-tag syntactic annotation system was developed, covering primary and secondary sentence parts while preserving national grammatical traditions. A key feature of the research is the treatment of Multi-Word Expressions (MWEs) using the BIO (Beginning-Inside-Outside) format, which ensures the syntactic integrity of complex predicates and modifiers. The resulting dataset consists of 9,025 sentences and 105,674 tokens, with 28.58% of tokens identified as part of MWE units. Statistical analysis reveals that this dataset provides a robust algorithmic foundation for training transformer-based architectures such as BERT, RoBERTa, and BiLSTM-CRF, overcoming the limitations of universal frameworks like Universal Dependencies for the Uzbek language