Intermittently, and for completely unknown reasons, users that havent done anything at all for literally months will upload some obviously fake "Lora" or "checkpoint" file with no description, fake screenshot (devoid of prompt or anything else), obviously nonsensical usage tips and absurd step or epoch count.
Sometimes it has a checkpoint file attached, other times its a zip file with screenshots inside and nothing else. I have never tried using these checkpoint files so I never discovered what they actually are. Does it have a virus built into it? Is it just a prank with that gives weird image results? Is it real and they just suck at data input?
My theory though? I think some of these users are not native english speakers and have the completely wrong impression that this website allows you to make your own Loras or checkpoints via some kind of algorithm. Suffice to say it does not. There are websites that do this but that's not something you can do here (at least not as of right now).
Yeah, some ppl really was doing it since CIvitAI popularization. Some ppl really don't know anything about CivitAI, for example - about posting images instead of "model page", and they just accidentally creating a whole model page for some their gens. It's okay, anyone can make a mistake.
I've seen a lot of chinese users just smacks their LoRA on models files "on table" and then disappears. Without any description, trigger words or something. Well, not sure what they trying to achieve.
There is really of lot of people that just uploads nonsensical things, as you sad - .txt or .zip files. Sometimes it contains advertisements and links to their paid things. sometimes it really has untagged datasets or complete datasets, but sometimes it can be infected files (.ckpt extensions, mostly) with obviously fake previews, that uses regular art from the internet. Not sure how fast these files got deleted by moderation, but i'm sure that CivitAI administration really need to make description for models or LoRAs a mandatory thing at least. And maybe add a requirement for at least one preview image to contain metadata with matching model, for new uploads (this would help to avoid a lot of tricky situations).