Basic

Comparing two runs honestly

A comparison is only fair when both runs faced the same set. If the set changed, the difference in score is partly the set and partly the model, and the two cannot be separated.

What you will see

Non-zero. A non-zero result is the signal to stop before reporting a delta.

About 1 min of reading

Try It

Freeze a set before comparing

1

Stamp the set with a hash

printf 'q1\nq2\nq3\n' > /tmp/set.txt; sha256sum /tmp/set.txt

Before you run it, say out loud what it will print. Then open the answer.

What it prints
One hash. That hash names the set for the rest of the comparison.
2

Change one case and re-hash

printf 'q1\nq2\nq3b\n' > /tmp/set2.txt; sha256sum /tmp/set2.txt

Before you run it, say out loud what it will print. Then open the answer.

What it prints
A different hash. Different hash means the scores are not comparable.
3

Compare the two hashes

cmp /tmp/set.txt /tmp/set2.txt; echo $?

Before you run it, say out loud what it will print. Then open the answer.

What it prints
Non-zero. A non-zero result is the signal to stop before reporting a delta.

Data Flow

Two runs to one fair delta
1 The set is frozen and stamped
↓
2 Model A runs the frozen set
↓
3 Model B runs the same frozen set
↓
4 The delta is reported with both rates
Why this works
Why the grader version is stamped

A grader that changed silently invalidates every past score. The version travels with the score.

What a tie means

A tie is a result, not a failure. It says the change is not measurable, which is grounds for shipping the cheaper thing.

Examples

The two things that must be identical

The case set and the grader. The model is the only part allowed to differ.

A comparison that moved the set

Adding three easy cases raises the score without the model improving. The win is an artefact of the set.

Why the set is frozen during a change

Freezing the set makes the number mean one thing: the model moved, or it did not.

Words this lesson uses

Each one is defined here on this site. Open the entry for the full definition and the other words it points at.

← Scoring a whole run
Previous Lesson
Scoring a whole run
Next Lesson
Eval set drift
Eval set drift →