Claims › Concept · How a verdict is made

Machine judge

Can a model stand in for a human rater?

A model can stand in for a human rater at a measurable rate of agreement; the rate is the claim, not the competence.

Where it sits

1 Free tag Bai a name finds, it does not vouch
2 Topic Evaluation · Architecture closed list, one per claim
3 Theme How a verdict is made derived from the topics
4 Concept Machine judge · 2 claims a note and links, or no build
5 Geometry 1.5 1.25 source, use, dissent

How its claims lean on each other

1 · Free tag2 · Topic3 · Theme4 · Concept5 · Geometryweight, 0.5 → 2The claimsBaiBaiEvaluationEvaluationArchitectureArchitectureHow a verdict is madeHow a verdict is madeWhat scale does not buyWhat scale does not buyVerdictVerdictMachine judgeMachine judgeAlso: Machine judge×1.5C021C021 · Asked to rank answers, a strong model agreed with human raters about as often as human raters agreed with each other, above eighty per cent.C023C023 · Harmlessness was trained by having the model criticise and revise its own answers against a written list of principles, instead of having people label every harmful output.

The claims

  1. C021

    “Once addressed, our results from 3K controlled expert votes and 3K crowdsourced human votes in the wild verify that GPT-4 judge match human evaluations at an agreement rate exceeding 80%, achieving the same level of human-human agreement (4.2, Table 4).”

    Asked to rank answers, a strong model agreed with human raters about as often as human raters agreed with each other, above eighty per cent. What was measured is a rate of agreement, which is not the same thing as competence to judge.

    Zheng et al. 2023 · §1 Introduction · contested · weight 1.5 · contests C019 · open C021

  2. C023

    “In this paper we develop a method we refer to as Constitutional AI (CAI), depicted in Figure 1, and use it to train a non-evasive and relatively harmless AI assistant, without any human feedback labels for harms. The method therefore improves upon, and partially replaces reinforcement learning from human feedback […] .”

    Harmlessness was trained by having the model criticise and revise its own answers against a written list of principles, instead of having people label every harmful output. Human judgement moves out of the labels and into the rules.

    Bai et al. 2022 · §1 Introduction · used · weight 1.25 · open C023

All 15 concepts →All the claims, one by one →