Hacker News new | ask | show | jobs
by astrange 855 days ago
This isn't true for an instruction-tuned model. They are designed so you actually do tell it what to do.
1 comments

Sure, but it's still a statistical model, it doesn't know what the instructions mean, it just does what those instructions statistically link to in the training data. It's not doing perfect forward logic and never will in this paradigm.
The fine tuning process isn't itself a statistical model, so that principle doesn't work on it. You beat the model into shape until it does what you want (DPO and varieties of that) and you can test that it's doing that.
Yeah but you're still beating up a statistical model that's gonna do statistical things.

Also we're talking about prompt engineering more than fine-tune