Large Language Models (LLMs) have demonstrated robust general-purpose task-solving capabilities across various domains, prompting researchers to investigate their potential for application in recommender system tasks. However, a significant semantic gap exists between LLMs and recommender systems: while LLMs excel at capturing language semantics, recommender systems rely on collaborative semantics, which limits the reasoning capability of LLMs for recommendations. Additionally, the majority of existing research adopts a generalized approach to user reviews, neglecting the possibility that users might express distinct sentiments toward different attributes in a single review. To address these challenges, we propose FADERec, a lightweight model combining fine-grained attribute-sentiment pairs distillation with collaborative signal learning. Specifically, FADERec employs a collaborative graph encoder to explicitly model user-item interaction patterns, extracting high-order collaborative signals from implicit feedback. These signals are then fused with attribute-sentiment pairs generated by Llama-2-7b. Extensive experiments on benchmark datasets demonstrate FADERec’s superiority, achieving state-of-the-art (SOTA) performance in both top-N and sequential recommendations.