PPFLOW is a target-aware peptide design method that uses torsional flow matching on torus manifolds to model the internal geometries of torsion angles in peptide structures. The method is accompanied by a new dataset called PPBench2024, which contains 15,593 high-quality protein-peptide pairs and is used to train deep learning models for structure-based peptide drug design. PPFLOW is evaluated on various tasks, including peptide drug generation, optimization, flexible re-docking, and side-chain packing, and shows state-of-the-art performance compared to baseline models. The method uses conditional flow matching to learn the distribution of torsion angles and global translation and orientation of peptides, and employs neural networks to approximate gradient fields for the flow matching process. The model is trained on a large dataset and is able to generate chemically valid peptides with high binding affinity and stability. The results show that PPFLOW outperforms other methods in terms of binding affinity, stability, validity, and novelty of generated peptides. The method also demonstrates strong performance in flexible re-docking and side-chain packing tasks, indicating its potential for structure modeling. The paper concludes that PPFLOW is a promising approach for target-aware peptide design and highlights the importance of large-scale datasets in training deep learning models for structure-based drug design.PPFLOW is a target-aware peptide design method that uses torsional flow matching on torus manifolds to model the internal geometries of torsion angles in peptide structures. The method is accompanied by a new dataset called PPBench2024, which contains 15,593 high-quality protein-peptide pairs and is used to train deep learning models for structure-based peptide drug design. PPFLOW is evaluated on various tasks, including peptide drug generation, optimization, flexible re-docking, and side-chain packing, and shows state-of-the-art performance compared to baseline models. The method uses conditional flow matching to learn the distribution of torsion angles and global translation and orientation of peptides, and employs neural networks to approximate gradient fields for the flow matching process. The model is trained on a large dataset and is able to generate chemically valid peptides with high binding affinity and stability. The results show that PPFLOW outperforms other methods in terms of binding affinity, stability, validity, and novelty of generated peptides. The method also demonstrates strong performance in flexible re-docking and side-chain packing tasks, indicating its potential for structure modeling. The paper concludes that PPFLOW is a promising approach for target-aware peptide design and highlights the importance of large-scale datasets in training deep learning models for structure-based drug design.