How do LLMs handle inconsistent training data?

drywall_dan

One aspect of LLM training that is confusing to me is how unsupervised learning works on data which is inconsistent. Suppose for example your training data consists of two people with different responses to the same question. How is this handled? It appears to me that a lot of data needs to be conditioned on the person saying it (or personality anyway).

Originally asked by FourierFlux on Stack Exchange (ai) (CC BY-SA 4.0).

uphilltech

Great question — answering as the UphillTech team:

What I have seen when I have inconsistent data (for text) is that multiple token have a high probability at certain points in the output: for example, if your question is like: How many fingers does a hand have? and you have 2 Answers in your training data, 4 and 5, then the output of the LLM could be something like 40% 4 and 45% 5. So you get the answer 4 about half of the time and 5 about half of the time. For other output I sometimes get inconsistent output but I did not use big models (more like SLM) for testing so output may be different for bigger models. The only point I have observe something like that in a real LLM is the seahorse emoji where many post on Reddit say it exists (it...

Full answer by joerg-de: https://ai.stackexchange.com/a/50250 (CC BY-SA 4.0).

Need it done for your business? UphillTech.