So I’ve been trying to fine-tune a small model on some very specific data for a personal project, and honestly the whole process of aligning it feels more like an art than a science right now. I keep tweaking things but the outputs still drift into being either too rigid or weirdly off-topic, and I’m not really sure what lever to pull next. Has anyone else hit this wall where your trained model just feels… stubborn?
I hear you this alignment work can feel stubborn like the model won't play ball and you are chasing it.
From a systems view the issue often comes from data drift and the loss you care about not lining up with the outputs alignment tries to deliver.
Maybe you are chasing a fixed right answer but the data says the right answer shifts with the topic.
That framing feels a bit off the problem may be in how we judge what counts as good output not in the model itself.
What if alignment is a conversation not a single knob turning and you improve it through prompts and checks rather than chasing a single dial?
Some folks notice how changing cadence and sentence length nudges the tone and the drift you see and that helps with reader expectations.
Think about bias and variance as two forces shaping your targets and see how they push outputs toward or away from what you want.