Weakly Supervised Video Anomaly Detection (WSVAD) trains models using only video-level labels to identify anomalies, which is a challenging task. Current methods rely on Multi-Instance Learning (MIL) and assume the segment with the largest feature magnitude in an anomalous video has a higher anomaly score than in a normal video. However, these methods often focus on the most significant anomalies (e.g., explosion flames), neglecting potential ones (e.g., non-violent robbery incidents), thus failing to recognize comprehensive anomalous events. Inspired by the human process of learning and applying knowledge, we propose an improved MIL strategy for WSVAD: a Vision-Language Instance Awareness Learning framework via Vision-Language Models (VLMs). The idea is to adapt the rich generic knowledge of VLMs to specific VAD application scenarios and thus uncover more encompassing anomaly event. Specifically, we first design a focus-glance video semantic extraction strategy based on VLMs for obtaining local-action and global-event semantics information in videos. Then, we effectively align knowledge understanding with visual facts and bridge the modality gap with a video-specific vision-language contrastive loss. Finally, we design a vision-langugage instance awareness learning pipeline with a multi-stage training process that bootstraps the video understanding strengths of the VLMs to different real-world VAD applications. Extensive experiments on the XD-Violence and UCF-Crime datasets demonstrate that our approach achieves state-of-the-art performance on the WSVAD task, and inference by textual modality alone still achieves promising performance.