The Digitised First Byurakan Survey (DFBS) contains low resolution optical spectra of more than 24'000'000 objects. However, the physical nature of majority of the sources is unknown. The size of the archive makes it impossible to examine the full archive by the common by-eye method.

Instead, machine learning positive/unlabelled classification method was applied to find distinct classes of objects, focusing to rare late type stars, such as M stars and carbon stars, findings a few new cases. These stars are good candidates for the exoplanet search.