I discuss the determination of the collinear fragmentation functions of light charged hadrons from a wide set of single-inclusive electron-positron annihilation cross section measurements. I focus on the recent effort made by the NNPDF collaboration, using a methodology based on Monte Carlo sampling for the representation of data uncertainties
and neural networks for parton parametrisation. I discuss the quality of the NNPDF determination, including its stability upon inclusion of higher-order corrections, and some of its implications in the investigation of the spin structure of the nucleon and in the description of the hadron spectra measured at colliders.