கKatralAI

Campus Katral · Lesson 6 · 11 min

Measure before you brag

In one sentence

A campus demo is not a grade. An eval set is how you prove the model did the job.

Nila

Nila

Campus mentor

AI is not a person. It is a next-word engine with a very good memory of patterns.

Read the idea, watch the picture, then try it with your hands.

Watch the idea

The concept

Case

Question + allowed sources

Score

Groundedness, not style

Gate

Invented citation is caught

A demo that works once is not a paper

Before you brag on GitHub, write cases: question, allowed sources, what good looks like, and when the system must refuse. Score groundedness, not vibe. That is how serious groups ship language features.

What belongs in a tiny eval set

Easy wins, hard retrieval, empty retrieval (must refuse), and a trap where a fluent lie looks legal. Twenty honest cases beat two hundred random chats you never scored.

Measure, then claim

Your README should say: on this set, unsupported answers dropped from X to Y. That sentence is hireable. 'It uses LangChain' is not. Evals turn a studio project into proof.

Worked example

The invented Act

The assistant cites 'Campus AI Act 2019' which does not exist. Style is perfect. Groundedness is zero. A gate that scores citations against the corpus fails it. A demo that only checks tone ships the lie.

Picture to keep

Cases → scores → fix

  1. 1Easy, hard, refuse
  2. 2Score the truth
  3. 3Fix, don't hide