This is a matter of expanding the definition, in this case Definition II.1.3 in SGA 4, which defines pretopologies.

By a “base” in this answer I mean what appears to be the most common definition: a collection of subsets of a fixed set $A$ such that any finite intersection of elements in the base is a union of elements in the base.

There are also multiple constructions of a site (i.e., a category with a coverage) from a base of a topological space $A$.
One can either (A) take the category of all open subsets of $A$,
or (B) the category whose objects are open subsets of $A$ that belong to the base.
For covering families of some object $V$, one can either take (a) those open covers of $V$ whose elements belong to the base, or (b) open covers of $V$ whose elements are given by the intersection of $V$ and some element of the base.
Altogether, there are four different options: A-a, A-b, B-a, B-b,
and only option A-b produces a pretopology in the sense of SGA 4.

Axiom PT0 for pretopologies says that any morphism in a covering family admits base changes.
Such base changes are always given by the corresponding intersection, provided that the intersection belongs to the category.
Thus, PT0 is satsfied for option A and not satisfied for option B.
If the base is closed under intersections, then PT0 is also satisfied for option B.

Axiom PT1 says that base changes of covering families are covering families.
In our case, the base change is given by the intersection with some open subset $V$.
Thus, PT1 is satisfied for option A-b and not satisfied for option A-a.
If the base is closed under intersections, then PT1 is also satisfied for options A-a, B-a, B-b.

Axiom PT2 says that covering families can be composed.
This is trivially true for the case under consideration.

Axiom PT3 says that the singleton family consisting of the identity map is a covering family, which is tautologically true in our case.

Thus, option A-b always gives a pretopology,
whereas options A-a, B-a, B-b give a pretopology
if and only if the base is closed under intersections of pairs.

In particular, open balls in a metric space as a coverage on the category of open subsets trivially form a pretopology using option A-b, and do not for a pretopology in any other option.