3 ms·
How does one even go about testing such a thing? Comparing it to Wikipedia articles? Even if it is factual does it spew the interpretations present in most Wi
by edgyquant 3y ago
How does one even go about testing such a thing? Comparing it to Wikipedia articles? Even if it is factual does it spew the interpretations present in most Wikipedia articles?
- commandlinefan 3y agoAfter having seen what professional "fact checkers" accept as fact (and reject as misinformation) makes me similarly skeptical.
- the_af 3y agoThey are not fact checking "truth" but whether the chatbot spouts "facts" supported by Wikipedia. This is objective and much easier to check than capital letter Truth. Consider that when discussing ideological or political articles, what is "true" becomes nebulous.
- the_af 3y agoThe abstract of the article explains what they mean: they mean the LLM does not hallucinate (so much) and provides facts based on Wikipedia. Absolute "truth" is not measured; rather, they measure how much the chatbot "sticks to the known facts" within Wikipedia. Since they are measuring this, presumably other chatbots and LLMs tend to hallucinate much more, providing "facts" not supported by their training data.
- jojobas 3y agoThe most irritating effect is that the LLM somehow guesses what you want it to respond. Human-in-the-loop training is imperfect.
- em-bee 3y agonow that is the interesting bit really. what makes the wikipedia based LLM hallucinate less? the only thing i can think of ad hoc is that wikipedia contains less conflicting or unclear information which helps to avoid the LLM getting confused. also the information is more organized, and it is clear which articles relate to each other. this would show what i think we already knew that LLMs can summarize the data they get but they can not evaluate or verify it.