Multivariate calibartion models developed for the radioisotope gauges designed for the measurement of the ash content in coal were investigated. The models were computed using the full spectra of the Am-241 g-rays scattered back and forward from the coal sample, as well as the spectra of X-ray fluorescence with backscattered radiation from Pu-238 source. The spectra of scattered g-rays were registered with the scintillation detector whereas for detection of X-rays the proportional counter was used. Calibration set contained 43 coal samples from four mines and the ash content ranged from 4 to 78%. The linear and non-linear partial least square (PLS) models were investigated and compared with multiple linear regression (MLR) models. It was found that the lowest value of root mean square error of crossvalidation was obtained for the non-linear PLS model in the case of the spectra of scattered g-rys and for the linear PLS model with preliminary data processing in the case of the XRF spectra.