The old Motorola App Note AN1304 discusses the use of coaxial baluns around 1GHz, to reduce Z mismatches of 100:1 to approximately 6:1.
also
......don't forget to check references listed in tha articles.
Going to 50ohms in the first place, may be part of the problem, as your active devices likely exhibit output impedances in the 10ohm range or less, to begin with. Matching earlier may help.
If galvanic isolation is required, a simpler 1:1 or 2:1 conventional balun transformer might be introduced additionally, in a convenient position. because required turns are low, you might consider winding shapes that are not normally seen in baluns - perhaps a single loop inside an isolated screen (~output winding).
With the app data aimed at achieving 50 ohms, it may be difficult to turn it around to make sense for low load values such as 0.5 ohms.
RL