[gigel82]

This is really bad, looks like the kind of output you'd get from a self-hosted GPT-2 instance: the article "contents" doesn't include any reference to the given topic, it's just repeated mumbo jumbo: https://www.paulgrahamai.com/recnOmtVDt9C4C8ss

[aubthread]

Yeah that's certainly a wonky result. What do you think would cause that? My thoughts would be (i) the lack of content specifically related to 'herding' or 'cats' or (ii) the ratio of completions (~200) to completion-length (~2k words each) of the fine-tune data. I'm new to this so appreciate any feedback.