Do threads score include testing on the first dataset? My processing timed out on the second one, but fit in time limit on the first one - by far not perfectly, of course, but to my taste - better than zero :)

Thanks for fixing that! Unfortunately I've made a typo in the output stage of "top" output, so instead of real category, all threads are assigned the same one (which is the category assigned to the very first processed article - typo results in reference to zero element instead of proper number), and realized there are other major problems. But nevertheless, now that it mostly works I feel much better :)

I forgot to put /models folder on the same level as tgnews binary. It requires models folder to be 1 level higher. Please either run it from nested folder, or put /models there

Thanks for the fix!
But there is a little problem. Languages are processed correctly, but everything else gives 0 results. It definitely is not what I wrote :) Can you please double check that you didn't break any logic when fixed output format? If you can send me fixed version, I would gladly check it myself and point on any possible problems there

Thanks for the review! Yes, noticed that - overall my algo uses a reference set of articles to categorize, and I've submitted terrible reference set (due to the lack of time). For example, in English I have 5 articles in "Science" (but relevant ones), while in Russian - 853 (95% clearly not science). I think I messed up score processing (should have calculate thresholds from actual distribution, while used static ones due to the lack of time), a few more days probably would make a big difference with my approach :) But right now it's nowhere close to relevant

I see... It's a pity, I have no idea what went wrong - on my setup everything works. Can you try running it on a different dataset? Maybe something smaller, I could have missed some memory errors associated with larger news arrays?