Step-function fit using leasqr

[Tweety]

Dear all,

I am trying to fit measured data using a step function, please see the 
attachments. I get an error "weighted residuals are not real". Ok, 
additionally to a conceptual glitch I might have bad initial values but 
I thought Marquardt Levenberg was supposed to converge even in that 
case. What am I missing?

I appreciate any suggestions. I might be on the wrong track with the 
model function and the initial values, so if there are any better 
approaches please let me know.

Why do I want to fit this data? Firstly, I need some kind of look-up for 
various x-values which might not coincide with measured x-values. 
Secondly, I need to get rid of the scatter in the end of the measurement 
for further processing. Is there perhaps a simpler way to achieve this?

Code for reference w/o attachment:

E = dlmread('data.txt');
x = E(:,1);
y = E(:,2);
q = [1; .25; .25];
function res = f(x,q)
  res = q(1) - exp((q(2).*x).^q(3));
end
[fr,q,cvg,iter_r] = leasqr(x,y,q,'f',.0001,40);

Any help would be appreciated. Thanks and kind regards, Jan

data.txt
Description: Text document

datafit.m
Description: Text Data