Arena (formerly LMArena)

What it is

Arena, previously called LMArena and before that Chatbot Arena, started as a research project by researchers from UC Berkeley and is now run by Arena Intelligence. You type a prompt, two anonymous models answer side by side, and you vote for the better answer, call it a tie or say both are bad. Only after voting do you see which models they were. The votes feed public leaderboards for text, coding, images, video and more.

Why it matters for AI trainer work

Side-by-side comparison, deciding which of two answers is better and why, is one of the most common AI trainer tasks. It produces the preference data used to tune models with human feedback. Arena lets you practise the core motion for free: read both answers fully, check facts, notice errors, and commit to a judgement.

How to practise properly

Voting quickly teaches little. Turn each comparison into a mini task:

1. Write a prompt from your own field, with a clear correct answer you can check.
2. Read both responses completely before forming a view.
3. Check every factual claim you can, and run any code.
4. Before voting, write two or three sentences of justification: which is better, the decisive reason, and the most serious flaw in the losing answer.
5. Rate each answer on a simple rubric (correct, complete, follows instructions, clear, safe) and see whether your rubric agrees with your gut.

Do this ten times and compare your written reasons. Vague reasons ("A is better") are the habit platforms reject; specific reasons ("B's second step uses the wrong formula, so its final number is wrong") are what they want.

Cautions

• Do not paste private, personal or confidential information into prompts.
• Leaderboards reflect many voters' preferences, which are not the same as a project's guideline.

Best for: anyone preparing for AI trainer work who wants free, realistic practice comparing model answers.