The reliability of spin polarization method results for atomic spin densities, obtained with several widely used Gaussian basis sets, is examined by comparison with the results of full configuration interaction (FCI) calculations. The spin densities obtained with these basis sets using the spin polarization model and some other methods disagree with the FCI treatment. Since the FCI wave function is exact for a given basis, it is not clear that the spin polarization model will be generally reliable. A large active space multiconfigurational (CASSCF) calculation is shown to be inadequate as an alternative to FCI treatment. The importance of accounting at least to some extent for excitations to all orbitals in the complete space of basis functions is illustrated by very slow convergence of CASSCF results with increasing size of active space. The FCI results reported here can be used as benchmarks to test various approaches to spin density calculation.