During a two-month period, Lszgz0521 contributed to bytedance-iaas/vllm by developing configurable CLS pooling in the ModernBertPooler, introducing support for multiple pooling strategies and robust error handling to prevent misconfigurations. This work, implemented in Python with PyTorch, improved model reliability and user experience for deep learning applications. In the embeddings-benchmark/mteb repository, Lszgz0521 expanded Japanese benchmark datasets by adding JaCWIR and JQaRA for retrieval and reranking tasks, while also addressing data quality issues in existing ANLP Journal datasets. Their efforts in benchmarking and dataset curation enhanced reproducibility and coverage for Japanese language information retrieval tasks.