We present a 'two-fluid' implementation of dust in smoothed particle hydrodynamics (SPH) in the test particle limit. The scheme is able to handle both short and long stopping times and reproduces the short friction time limit, which is not properly handled in other implementations. We apply novel tests to verify its accuracy and limitations, including multi-dimensional tests that have not been previously applied to the drag-coupled dust problem and which are particularly relevant to self-gravitating protoplanetary discs. Our tests demonstrate several key requirements for accurate simulations of gas-dust mixtures. Firstly, in standard SPH particle jitter can degrade the dust solution, even when the gas density is well reproduced. The use of integral gradients, a Wendland kernel and a large number of neighbours can control this, albeit at a greater computational cost. Secondly, when it is necessary to limit the artificial viscosity we recommend using the Cullen & Dehnen (2010) switch, since the alternative, using α ∼ 0.1, can generate a large velocity noise up to σ v 0.3c s in the dust particles. Thirdly, we find that an accurate dust density estimate requires > 400 neighbours, since, unlike the gas, the dust particles do not feel regularization forces. This density noise applies to all particlebased two-fluid implementations of dust, irrespective of the hydro solver and could lead to numerically induced fragmentation. Although our tests show accurate dusty gas simulations are possible, care must be taken to minimize the contribution from numerical noise.