Fitting data to beta distribution

What is the actual likelihood function being maximized when fitting data to the beta distribution using 'betafit', specifically when the data are censored (i.e. when the data contain 0's and/or 1's)? Also, how is the gradient of this likelihood being approximated?