A health analogy for training LLMs
A post likens training data to food, regular model evaluations to health checkups, and eating lots of sugar to reward hacking that raises dopamine or metrics.
TLDR
A user compares caring for an LLM with caring for your body: datasets feed the model, and regular evaluations resemble health checks. They liken eating lots of sugar to reward hacking and argue people should care as much about their health as they do about LLMs.
Combined views
1.7K
1 Source, first seen 3h ago
likes