@nprofile1q... there are some groups working toward this kind of thing, if you're not already aware (e.g. https://www.writefutureslab.com/little-free-language-models/metaphor-workshops/). personally i would love to have a model like this for uses like classification, clustering, search, etc., with the knowledge that the model will still encode the underlying semantics and positionalities of the dataset. but i think the dangers/drawbacks of chatbot-style interaction ("psychosis," anchoring bias, etc) would still apply even to a small model