If modifying the input sigma doesn't help, it's likely that for your call eigs(A, B, sigma, ...) the matrices are such that A - sigma*B is always singular for any sigma.
This would mean there is at least one vector v for which A*v = 0 and B*v = 0, meaning this is an "eigenvector" for every possible eigenvalue sigma. That means the eigenvalue problem is ill-posed.
This is not so uncommon for matrices coming from FEM, it basically means the matrix size is larger than than the degrees of freedom in the underlying problem.
I'm not an expert on FEM, but for example, if we apply Dirichlet conditions to the first basis, this likely means that the first row and column of both A and B is all zero. That would lead to exactly the issue above. Removing the first row and column of both A and B would result in a smaller and well-posed problem.
Possibly freefem++ provides some options that would cut out these additional degrees of freedom? I don't know the software well enough to help with this.
