ContentsThe library

Training a Model From Scratch

Three Ways a Run Breaks, and What Each One Looks Like

Last timeReading a Run While It Happens

A loss spike, a divergence and a dead plateau look alike for about ten steps and need completely different responses. Here is how to tell them apart.

Three shapes

Runs break in a small number of ways, and the shapes are distinct enough to name.

FIG 1Three curves
0.001.252.503.755.000.05.010.015.020.0training progress, in thousands of steps
healthydivergingdead
The diverging run looks healthy for the first few thousand steps. The dead run never looked like anything else.

A spike is a sudden jump in the loss followed by a return to roughly where

it was. It is not on that chart because it is a single point: the loss goes from

2.2 to 7 for one step and is back at 2.4 a hundred steps later.

A divergence is a jump that does not come back. The loss climbs, or goes to

infinity, or becomes the special value that means "not a number" and stays there

for the rest of the run.

A dead plateau is a loss that is flat from the beginning, usually at a

suspiciously specific value, and never moves.

The responses are different enough that getting the diagnosis wrong is

expensive. A spike treated as a divergence throws away a working run. A

divergence treated as a spike wastes however many hours it takes to notice.

A spike, and what it costs to undo

A spike is almost always one batch. Something in it produces an unusually large

gradient, the step lands somewhere bad, and the next few hundred steps are the

model finding its way back.

Most spikes resolve themselves, and the right response to most of them is

nothing. The decision worth making in advance is at what point you stop waiting,

and the answer comes out of your checkpointing interval, because rolling back

costs exactly the work since the last checkpoint.

The lesson stops here

6 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 contents

This 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

The rest of this course

  1. 01Four Lines, and Genuinely All of It
  2. 02The Four Hops Between Storage and the Stepopening only
  3. 03The Starting Scale, and Why It Decides Everythingopening only
  4. 04One Number, and Why It Cannot Be One Numberopening only
  5. 05What a Bigger Batch Actually Buysopening only
  6. 06The Four Numbers Worth Watchingopening only
  7. 07Three Ways a Run Breaks, and What Each One Looks Likeyou are here
  8. 08Nobody Trains to Convergenceopening only

Read alongside