I see. Indeed we can try to add a practical guide to the Wiki in order to find the appropriate settings for LANGEVIN_GAMMA, LANGEVIN_GAMMA_L and PMASS. There are probably many ways one can arrive at reosanable settings. By reasonable seetings I mean that the instantanious temperature and instantanious pressure fluctuations are within the expected range according to the system size and stiffness and there should be no systematic drifts. There are some statistical tests you can find in literature, but I am not an expert in that area and it is rather a scientific question, I'd say.
Step 1: Start with an NVT ensemble to reduce the adjustable parameters. Thermalize the structure and plot the temperature as a function of MD steps. If you see a drift, you may need to decrease POTIM, or choose tighter settings for the ab initio calculation (ENCUT, PREC, etc). If you see basically no fluctuations, the ions are not coupling to the thermal bath. Hence increase the coupling by reducing the dampning (set a smaller LANGEVIN_GAMMA).
Step 2: Release the lattice degrees of freedom by setting ISIF=3 and choosing LANGEVIN_GAMMA_L and PMASS. The lattice vectors are now entering as degrees of freedom with there own ficticious mass (the lighter the more dynamic) and dampning (larger LANGEVIN_GAMMA_L yields less fluctuation). Usually you only adjust LANGEVIN_GAMMA_L and not PMASS, but thats a matter of taste. Again plot the pressure as a function of MD steps and find the best LANGEVIN_GAMMA_L to get no drifts and reaosanable fluctuations.
Mind that the goal is not top have no fluctuations. That goes along with bad ergodicity, i.e., you will have difficulty to visit all of the phase space.
Can you use this to setup a calculation and let me know if you obtain good MD trajectories?