The model doesn't know about itself
#17
by sakazakiMGJ - opened
I noticed with all the small Qwen models (4B, VL 4B, 4B instruct), they think that they are other bigger version. They think they are 70B versions or other bigger version and that they are in the cloud. They can even get in an infinity loop trying to think that they are not that parameter size or that they don't even exist.
They are probably distilled from the bigger ones' outputs.
Just like idiots in real life lol