Random Monsters Map Patcher – diverse neutral armies.

Calculating weights is too complex. A large number of tests by the AI are needed in various combinations. Even if we use pairs, there will be many. Although, we can exclude some combinations as ineffective and assign them a weight of 0.

Perhaps, we'll have to switch to a rating system, where combinations will be ranked according to a table of their effectiveness in player usage (again, a dumb AI could ruin everything).
In this case, the rating will only reflect the priority, and it will not be clear how much better 1st place is than 2nd place, for example.

The third system is a point-based one.
We assign expert ratings from 0 to 10 to all pairs. From these ratings, we can, in principle, find the weight, but then the number of experts must be greater, and they should be more knowledgeable about AI actions than about playing the game themselves (due to the AI's inability to play units well).
It is permissible to give overpowered combinations more than 10 points.
The third system turns out to be a kind of average between the first and second. This might work. Here, we can conditionally sum the points as in the weights.

Added 2 minutes ago

Otherwise, the same additional units will correspond to some creatures, and they will be repeated frequently. What exactly is the loss of meaning in the mod?

Repetitions will be excluded.

Excluding "evil" units. If 10 mages appear in stacks of 2 each randomly, that's even worse than just having 10...
And the mod should have more meaning.