With the emergence of large language models (LLMs), generative artificial intelligence has gained traction in lyrics creation. Compared with pre-trained language models (PLMs), LLMs can designate the detailed structure and semantic content of lyrics through the capability of Instruction-following. Some research explicitly model the structural and semantic features of lyrics but simplify the definition of lyrics features.To address these issues, we propose a lyrics generation model named Okashi that models structural and semantic features of lyrics simultaneously. Additionally, we incorporate lyric structural features into the analysis of semantic features to ensure their underlying consistency. We construct three Japanese lyrics datasets and annotate of structural and semantic labels automatically through carefully designed prompt templates, avoiding costly human annotation. Experiments demonstrate that our model outperforms baseline models in terms of lyric quality and diversity. Okashi can generate lyrics with clear structure, high text quality, and conformity to human lyrics writing principles, showcasing the effectiveness of our structural and semantic feature modeling.