Depends on model tuning. Basically, you can tune a model to hallucinate less, or to write more human-like, but not really both at the same time, at least not for a model you could expect most users to run locally. For this sort of application (summarizing text), you’d tune heavily against hallucination, because ideally your bullet points are going to mostly be made up of direct paraphrase of article text, with a very limited need for fluid writing or anything even vaguely creative.
Basically, you can tune a model to hallucinate less
You can tune it to hallucinate more, you can’t tune it to not hallucinate at all, and that’s what matters. You need it to be “not at all” if you want it to be useful, otherwise you can never be sure that it’s not lying, and you can’t check for lies other than reading the article, which defies the whole purpose of it.
Depends on model tuning. Basically, you can tune a model to hallucinate less, or to write more human-like, but not really both at the same time, at least not for a model you could expect most users to run locally. For this sort of application (summarizing text), you’d tune heavily against hallucination, because ideally your bullet points are going to mostly be made up of direct paraphrase of article text, with a very limited need for fluid writing or anything even vaguely creative.
You can tune it to hallucinate more, you can’t tune it to not hallucinate at all, and that’s what matters. You need it to be “not at all” if you want it to be useful, otherwise you can never be sure that it’s not lying, and you can’t check for lies other than reading the article, which defies the whole purpose of it.