multinomial logistic regression (mnrfit) with fisher's iris data - MATLAB issue?

So i have been writing my own code to perform a multinomial logistic regression. I have been using MATLABs mnrfit to check that the MLE's agree. All seems to work well. I can simulate data, fit using both mnrfit and my code and the estimates agree. However when i use fisheriris data, used in MATLAB's example (<http://www.mathworks.com/help/stats/mnrfit.html>) our estimates do not agree. In fact if i look at the value of the likelihood using mnrfit's parameter estimates it is no where near its maximum. I can not understand why?

In addition when i reconstruct the probabilities using the parameter estimates from my code, i find that for flower setosa all of the probabilities are 1. Not so under mnrfit. And since this data set is commonly used to demonstrate discriminant analysis this result seems more likely. I also wonder why the probit example given in glmfit (<http://www.mathworks.com/help/stats/glmfit.html?searchHighlight=glmfit>) uses only the latter 2 species in fisheriris

Can anyone comment on this? Anyone from Mathworks care to comment? I am quite confused. Thanks