Glossary
Plain-English definitions of 19 terms that show up in AI training and evaluation job listings - what a task actually is, with a real example, before you apply to it.
The step-by-step reasoning an AI model shows before arriving at a final answer.
Labeling raw data (text, images, audio, video) so it can be used to train or evaluate an AI model.
Someone with real professional experience in a specific field (medicine, law, engineering, finance, etc.), hired specifically for that expertise rather than AI knowledge.
Further training an already-trained AI model on a narrower, more specific dataset to specialize its behavior.
A verified, correct reference answer that an AI's output is compared against.
The verified, correct answer or label that a model's output (or an annotator's label) is measured against.
When an AI model states something false or fabricated as if it were fact.
Any AI system design where a human reviews, corrects, or approves the model's output as a routine part of the process.
A measure of how consistently different human reviewers label or score the same piece of data.
Judging the quality, accuracy or safety of an AI model's output, usually against a rubric written by the client.
Pay structured as a fixed amount per completed task rather than an hourly wage.
Ordering two or more AI-generated responses from best to worst, the raw data that RLHF and reward models are built from.
Designing and refining the exact wording given to an AI model to reliably get the output you want.
Deliberately trying to make an AI model fail, break its rules, or produce harmful output, so the failure can be fixed before release.
A separate model trained on human preference data, used to score how "good" a candidate AI response is during further training.
A training method where humans rank or rate AI outputs, and those judgments are used to steer the model toward responses people prefer.
Scoring an AI's response against a specific, predefined checklist of criteria rather than a general "is this good" judgment.
Training a model directly on example input-output pairs written or corrected by humans, rather than on ranked preferences.
Labeling text or model output as harmful, offensive, or policy-violating so a model can learn to avoid or catch it.