AI training glossary
The words you'll meet in listings, guidelines and worker forums, defined plainly. 32 terms.
- AI interview
- A screening conversation run by an AI interviewer, used by Mercor and micro1. It asks about your background and field, records a transcript and scores it. Read more →
- AI trainer
- A contractor who writes, checks or grades AI model outputs so the model can learn from human judgment. Also called AI tutor, data trainer or human data expert. Read more →
- Annotation
- Adding labels to data: tagging images, marking entities in text, transcribing audio. The broad category AI training work grew out of.
- Assessment
- A test that looks like the real work, used to decide whether you get access to projects. Often unpaid at the starter level, sometimes paid for expert projects.
- Attempter
- Platform slang for the person doing a task, as opposed to the reviewer who checks it.
- Benchmark
- A fixed set of questions used to measure model performance. Experts are often paid to write benchmark questions the current model gets wrong.
- Chain of thought
- The step-by-step reasoning a model writes before its final answer. Some tasks ask you to grade or write the reasoning, not just the answer.
- Contractor (1099)
- How almost every platform pays you: as a self-employed independent contractor, with no tax withheld. US platforms send Form 1099-NEC once they pay you $2,000 or more in a year (from 2026). Read more →
- DPO
- Direct preference optimization: a training method that learns from the same human preference comparisons as RLHF but without a separate reward model. Read more →
- EQ (empty queue)
- Worker slang for having no tasks available. Common between projects on Outlier and similar platforms.
- Evaluation (evals)
- Testing how well a model performs. Eval work means grading model answers rather than producing training data.
- Expert track
- A higher-paid project category that requires a degree, license or demonstrated professional skill in a field such as medicine, law or finance.
- Fine-tuning
- Further training of an existing model on a smaller, targeted dataset, such as answers written by experts.
- Golden task
- A task with a known correct answer, mixed in with real work to check quality. Missing golden tasks can lower your score or remove you from a project.
- Guidelines
- The project's instructions document. On most projects, following it exactly matters more than your own view of what's best.
- Hallucination
- A confident statement from a model that is false or made up. Catching hallucinations is a core part of fact-checking tasks.
- Human data
- The industry term for data created or judged by people for AI training: rankings, written answers, rubric scores and so on.
- Multimodal
- Models and tasks that combine text with images, audio or video.
- Onboarding
- The instructions, videos and quizzes you complete before starting a project. Whether onboarding time is paid varies by platform and project.
- Preference ranking
- Comparing two or more model answers and choosing the better one, usually with a reason. The basic unit of RLHF data. Read more →
- Prompt
- The question or instruction given to a model. Prompt-writing tasks pay you to write realistic, difficult prompts.
- Red teaming
- Deliberately trying to make a model produce harmful, false or rule-breaking output, and documenting how, so labs can fix it.
- Referral bonus
- A payment a platform makes to someone who refers a new worker, usually after the new worker logs a set number of paid hours. Read more →
- Reviewer
- A trainer, usually experienced, who checks other trainers' tasks. Reviewer roles often pay more.
- Reward model
- A model trained on human rankings to predict which answer a person would prefer. RLHF then trains the main model to score well with it.
- RLHF
- Reinforcement learning from human feedback: training a model with people's judgments about which outputs are better. Read more →
- Rubric
- A checklist of criteria for grading an answer, such as accuracy, completeness, safety and format. Expert projects increasingly use detailed rubrics.
- SFT
- Supervised fine-tuning: training a model on example answers written by people. Writing 'ideal responses' is SFT work.
- Starter assessment
- DataAnnotation's entry test. No rejection email is sent; if you pass, projects appear in your dashboard. Read more →
- Task pay vs hourly pay
- Some projects pay a fixed amount per task instead of per hour. We only convert to hourly when the listing gives a real hourly figure. Read more →
- Time cap
- The maximum paid time allowed for one task. Going over usually means the extra time isn't paid.
- Work trial
- A short paid or unpaid sample project before a contract, used by some expert listings.