ABSINSTRUCT is a framework designed to enhance the abstraction ability of large language models (LLMs) through instruction tuning with plausibility estimation. The framework generates detailed explanation traces for each example, helping LLMs understand the underlying reasoning behind abstract concepts. It also uses a plausibility estimator to select instructions that align with the knowledge of pre-trained models, ensuring the data is relevant and effective. The framework combines abstraction instructions with general-domain instructions to create a hybrid dataset, which is then used to fine-tune LLMs. Extensive experiments show that ABSINSTRUCT significantly improves LLMs' abstraction ability, outperforming existing alignment methods by up to 10%. The framework also maintains the LLMs' general instruction-following capabilities. The approach includes data collection, explanation trace generation, plausibility estimation, and filtering to ensure high-quality, diverse, and relevant data. Evaluation on multiple datasets demonstrates the effectiveness of ABSINSTRUCT in enhancing abstraction ability without compromising general performance. The framework is evaluated on various tasks, including the ABSPYRAMID benchmark and out-of-domain datasets, showing strong generalization and robustness. The results highlight the importance of abstraction knowledge in NLP tasks and demonstrate the potential of ABSINSTRUCT in improving LLMs' capabilities in this area.ABSINSTRUCT is a framework designed to enhance the abstraction ability of large language models (LLMs) through instruction tuning with plausibility estimation. The framework generates detailed explanation traces for each example, helping LLMs understand the underlying reasoning behind abstract concepts. It also uses a plausibility estimator to select instructions that align with the knowledge of pre-trained models, ensuring the data is relevant and effective. The framework combines abstraction instructions with general-domain instructions to create a hybrid dataset, which is then used to fine-tune LLMs. Extensive experiments show that ABSINSTRUCT significantly improves LLMs' abstraction ability, outperforming existing alignment methods by up to 10%. The framework also maintains the LLMs' general instruction-following capabilities. The approach includes data collection, explanation trace generation, plausibility estimation, and filtering to ensure high-quality, diverse, and relevant data. Evaluation on multiple datasets demonstrates the effectiveness of ABSINSTRUCT in enhancing abstraction ability without compromising general performance. The framework is evaluated on various tasks, including the ABSPYRAMID benchmark and out-of-domain datasets, showing strong generalization and robustness. The results highlight the importance of abstraction knowledge in NLP tasks and demonstrate the potential of ABSINSTRUCT in improving LLMs' capabilities in this area.