The Wrong Answers Are Where the Teaching Is
Last timeTaking Weights Out Entirely
A small model trained on a large one's full output distribution learns more than the same model trained on the original labels. Here is the objective and why it works.
Every technique so far has taken a large model and described it more cheaply.
This one builds a different, smaller model and teaches it to behave like the
large one. The result is genuinely small: fewer layers, narrower, less state
per request, faster arithmetic. All three budgets from the first lesson fall at
once.
The interesting part is not the smaller architecture, which anybody can
specify. It is that training the small model on the large one's outputs works
considerably better than training the same small model on the original data,
and the reason is worth understanding because it tells you what to match.
What a label leaves out
Take one position in a sentence: "the capital of Australia is ___". The
training data says the answer is Canberra. That is the whole of what the label
communicates: one token is right and the other fifty thousand are wrong, with
no distinction among them.
The teacher, asked the same question, produces a probability for every token in
its vocabulary.
That ranking is the thing being transferred. It says Sydney and Melbourne are
the kind of thing that goes here; it says the answer is a city; it says
punctuation and verbs are not candidates. The label says none of it, and a
small model learning from labels alone has to rediscover all of it from scratch
with far less capacity than the teacher had.
The lesson stops here
4 more paragraphs to go
You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.
The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.
See the planThe contentsThis is the reading half
Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.
The contents