ContentsThe library

Reading a Benchmark Honestly

The Top of the Table Is Not Where the Best System Is

Last timeIs the Gap Real

Take enough noisy measurements and keep the largest one and you have measured luck. Leaderboards do this by design, and a test set reused for a year stops being a test set.

A leaderboard has six hundred entries. The top one scores 92.8. The question

worth asking is not whether that system is good but what the number 92.8 is a

measurement of, and the answer involves the other five hundred and ninety nine

entries.

Why the top of a table is inflated

Every entry's score is its true quality plus a draw from the uncertainty of the

previous lessons. Most entries draw something small. A few draw unusually well.

When you sort the table and look at the top, you are looking at whichever entry

had the most favourable combination of being good and being lucky, and you

cannot tell from the number which part contributed more.

FIG 1How far above the truth the winner sits
the measured score of the top entry
the true quality of that entry
the standard error of a single measurement
how many entries were measured
This is the expected largest of many draws, and it says two useful things. The inflation grows only slowly with the size of the field, through a logarithm, so a field of six hundred is not much worse than a field of sixty. But it scales directly with the error on a single measurement, so a small question set makes the whole table untrustworthy at once.
FIG 2Expected inflation of the top entry as the field grows
0.003.757.5011.2515.002.076.5151.0225.5300.0entries on the leaderboard
measured on a hundred questionsmeasured on four hundred questions
Both curves rise steeply at first and then crawl. Going from two entries to twenty costs most of the damage; going from twenty to three hundred adds a few points more. The gap between the curves is the point: the only lever that really helps is making each individual measurement sharper.
FIG 3Inflation of the top entry in percentage points
100 questions400 questions1600 questions
4 entries6.23.01.5
16 entries8.74.22.1
64 entries10.75.22.6
256 entries12.36.03.0
The marked cell is a common arrangement and a bad one: a widely used set of a hundred questions with a couple of hundred published results against it. The leader of that table is expected to be about twelve points above its own true quality, which is larger than the entire spread of the top twenty entries.

Notice what this does not say. It does not say the top entry is bad, and it does

not say the ranking is random. Quality still matters and the good systems still

cluster near the top. What it says is that the number attached to the winner is

not a measurement of the winner, and that the difference between first and fifth

place on a crowded table is usually smaller than the noise that sorted them.

The lesson stops here

5 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. 01The Number Describes the Questions as Much as the Model
  2. 02How Far the Number Could Have Landed From Where It Didopening only
  3. 03Only the Questions They Disagree About Carry Any Informationopening only
  4. 04The Top of the Table Is Not Where the Best System Isyou are here
  5. 05A Model That Has Seen the Exam Is Not Being Examinedopening only
  6. 06The Name on the Box Is a Claim, Not a Descriptionopening only
  7. 07Everything Between the Question and the Numberopening only
  8. 08Say the Number, the Range, the Conditions and the Limitsopening only

Read alongside