Skip to lesson

Path: How post-training works

Evaluation closes the loop

Use held-out tasks and explicit criteria to check whether a change helped without creating new failures.

Starting level
Some basics
Time
20 minutes
Public lesson guide
EN + हिंदी

You will learn to

Design a small held-out evaluation that checks improvement and regression.

Why this matters

A change can improve one metric while making accuracy, safety, or a subgroup worse.

Three things to remember

  • Keep evaluation tasks separate from development examples.
  • Measure the intended benefit and likely regressions.
  • Inspect important languages, domains, and user groups.

See one example

Task

A team changes a model to make answers shorter.

Weak approach

Check one prompt used during development.

Stronger approach

Compare a held-out set before and after for length, correctness, safety, and important subgroups.

Why the stronger approach works

A multidimensional held-out test can reveal both the intended gain and hidden regressions.

Make something yourself

Create a five-task evaluation plan with one success metric, one safety check, and one subgroup check.

A useful starting pointState what remains fixed between the before and after comparison.
Practise this lesson

Read further

The source is optional. The complete teaching and practice are available inside AI2Bharat.

Read the primary or official source: OpenAI HealthBench