[rytill]

LLMs learn a distribution during pre-training, not only an average.

Then, by giving them context or by post-training, you can make them sample non-average parts of the distribution they learned.

[OtomotO]

> Then, by giving them context or by post-training, you can make them sample non-average parts of the distribution they learned.

How do you derive that something is "below average" or "average" or "above average"?

[rytill]

Well, it’s up to the user or post-trainer of the LLM what they believe to be above average. Then they can design around that.

In the case of real world LLMs and post-training, what is above average is defined roughly as: labeled good by expert humans, and scoring high on RL environments related to coding like debugging, passing tests, or running efficiently and verifiably correctly.

[nextaccountic]

> How do you derive that something is "below average" or "average" or "above average"?

One technique is RLHF: have an human expert assess it.

[OtomotO]

Mhm, I just wonder how many samples they get and how much time they have to come to the conclusion.

Like a short example is easier to grade, but not in the same ballpark as a whole codebase.

[u8080]

>How do you derive that something is "below average" or "average" or "above average"?

How do you? I mean, that was your point basis.