3 ms·
That's why you fine tune models, right? To make sure the undesirable things in their data set don't lead to undesirable output.
by staticman2 3y ago
That's why you fine tune models, right? To make sure the undesirable things in their data set don't lead to undesirable output.
- xeckr 3y agoFine-tuning will never get rid of all biases introduced by the training data, some of which are subtle enough that you won't even know that you should apply fine-tuning techniques to try to mitigate them. For example, did you know that there is a trend for physics textbooks to be dedicated to Xi Jinping? If you didn't know this, because who has time to sift through millions of books to find undesirable biases, isn't it clear that your LLM will develop subtle but favourable biases towards the CCP?