The bernardi_COMPOSITES.tar file includes the 925 composite spectra 
(spec_"ID".fit files).

Each .fit file is a matrix of [3000,4]:

   FLUX            FLOAT     Array[3000]
   FLUXERR         FLOAT     Array[3000]
   FLUXMEASERR     FLOAT     Array[3000]
   WAVELENGTH      FLOAT     Array[3000]

 -- the first row [*,0] gives the co-added spectrum 
    flux density in [10^(-17) erg/s/cm^2/A] normalized
    to the mean flux of the galaxies used to make the 
    composite spectrum; 
 -- the second row [*,1] gives the observed errors, the rms scatter
    rms_obs/sqrt(N) (i.e. the observed dispersion of the individual 
    galaxy spectra around the average flux), of the normalized co-added 
    spectrum flux density. This is the sum in quadrature of the measurement 
    error and of the intrinsic scatter of the different galaxy spectra
    which make the composite: rms_obs=sqrt(err_meas^2 + rms_int^2).
    However for some pixels the rms_obs is smaller than the err_meas
    (due to small number statistics). For such pixels we substitute
    rms_obs with err_meas (note that this is an underestimate of
    the true observed error).
 -- the third row [*,2] gives the mesurement error err_meas
    (one can check where we substitute rms_obs with err_meas by
    comparing the second [*,1] and third [*,2] rows). 
 -- the fourth row [*,3] gives the wavelength in Angstroms.

The "ID" number corresponding to the spec_"ID".fit file is given in 
column 1 of Table 1-4 and 7-8 in Bernardi et al. (and in the electronic 
version of those Tables -- see bernardi_TABLES.tar.gz).