Adding fake subdomains to a low-resource domain.

[Martovark]

I sometimes encounter a situation where there are very few examples of subdomains for a certain domain, for example, there is only www.some_domain.com or subdomain.com and there is no more data.

Two simple ideas arise:

• Before tokenization, add fake subdomains that will provide additional information for the model and possibly improve search results. For example [git, gitlab, ...], or [ftp, mail, ...]. It seems like a difficult question, which subdomains to choose. Perhaps it is worth considering the distribution of the training data.
• Enable the option to automatically restart subdomain scanning using information from newly found subdomains. Sure, you could get a new list of domains, paste it into example_input.txt, and restart, but that doesn't seem very convenient. You can continue searching in a loop until no new domains appear.

[klaasmeinke]

Thanks for your questions @Martovark! And sorry for taking a while to get back to you.

• Regarding the first question: we actually get our training data from the certificate transparency log so the GPT model is trained on millions of real life combinations of subdomains.
• Regarding the second question: in release v0.4.0 (August 2025) we actually added recursion, so the tool does now find new subdomains, and then rerun the model using those subdomains as inputs.