Proteins can form biomolecular condensates in a process called liquid-liquid phase separation (LLPS). These condensates can be formed by many proteins with different roles, either leading its formation (drivers) or being participants (clients). Many databases are developed to collect LLPS proteins but the level of annotations widely varies between them, preventing fair benchmarks and specific predictive tools. The datasets here deposited are generated by rationally filtering the most relevant LLPS resources that eventually integrate confident annotations of client, driver and negative proteins in the LLPS framework.