AI training glossary
The words you meet as an AI trainer, from RLHF and red teaming to gold tasks and empty queues, each defined in plain English.
How models are trained
- RLHF (Reinforcement Learning from Human Feedback)
- A training method where people compare or rate model answers, and those judgments teach the model which answers to prefer. Most AI trainer work on rating and ranking tasks feeds RLHF.
- More on RLHFAI Safety & Evaluation roles
- SFT (Supervised Fine-Tuning)
- Training a model on example prompts paired with ideal answers written by people. Writing those answers is a common task for domain experts.
- More on SFT
- Preference data
- Pairs or sets of model answers with a human judgment of which one is better and why. It is the raw material for RLHF and similar methods.
- More on Preference data
- RLAIF (Reinforcement Learning from AI Feedback)
- A variant of RLHF where another model gives the feedback. People still write the rules it follows and check its judgments.
- More on RLAIF
- DPO (Direct Preference Optimization)
- A way to train on preference data directly, without a separate reward model. For the person rating answers, the task looks the same as RLHF.
- More on DPO
- Reward model
- A model trained on human preference data to score answers, which then guides the main model during training.
- More on Reward model
- RL environment
- A task setup, such as a coding sandbox or a simulated website, where a model attempts work and is scored automatically. Experts are hired to design tasks and grading rules for them.
- More on RL environmentSoftware & IT roles
- Synthetic data
- Training data generated by a model rather than written by a person. Human reviewers check and correct it before it is used.
- More on Synthetic data
- Human-in-the-loop (HITL)
- Any process where a person reviews, corrects or approves machine output before it counts. Nearly all AI training work is a form of it.
- More on Human-in-the-loop
Task types
- Data annotation
- Adding labels to raw data, such as tagging objects in an image, marking entities in text or timestamping speech, so a model can learn from it.
- More on Data annotationBrowse open roles
- Data labeling
- Another name for data annotation, usually for simpler, high-volume tasks with a fixed set of labels.
- More on Data labelingBrowse open roles
- Model evaluation
- Judging model answers against a rubric for accuracy, helpfulness, safety and style. Specialist evaluation is among the best paid work in this field.
- More on Model evaluationAI Safety & Evaluation roles
- Response ranking
- Ordering two or more model answers to the same prompt from best to worst, usually with a short written reason.
- More on Response ranking
- Prompt writing
- Writing questions or instructions designed to test a model, often hard ones in your own field that the model is likely to get wrong.
- More on Prompt writing
- Red teaming
- Deliberately trying to make a model produce harmful, false or policy-breaking output, so the weaknesses can be fixed before release.
- More on Red teamingAI Safety & Evaluation roles
- Fact-checking
- Verifying each claim in a model answer against reliable sources and marking what is wrong or unsupported.
- More on Fact-checking
- Transcription
- Writing out what is said in audio or video, often with speaker labels and timestamps, to train speech recognition.
- More on TranscriptionLanguages & Linguistics roles
- Bounding box
- A rectangle drawn around an object in an image or video frame, labeled with what the object is. A basic computer vision task.
- More on Bounding boxVideo, Image & Design roles
- Semantic segmentation
- Labeling every pixel of an image with the class it belongs to, such as road, car or person. Slower and more precise than bounding boxes.
- More on Semantic segmentationVideo, Image & Design roles
- Named entity recognition (NER)
- Marking names of people, places, organisations, dates and similar items in text.
- More on Named entity recognition
- Sentiment labeling
- Marking whether a piece of text is positive, negative or neutral, sometimes with finer emotions.
- More on Sentiment labeling
- Code review task
- Checking model-written code for bugs, security problems and style, then rating it or writing a corrected version.
- More on Code review taskSoftware & IT roles
- Golden answer
- A model answer written by an expert to be the reference everything else is graded against. Also called a gold standard response.
- More on Golden answer
Quality and review
- Rubric
- The written scoring guide for a task: what to check, how to score each point and what counts as a failure.
- More on Rubric
- Guidelines
- The project instructions for a task. Following them exactly matters more than personal judgment, and reviewers score against them.
- More on Guidelines
- Gold task
- A hidden test task with a known answer, mixed into normal work to measure accuracy. Missing too many can remove you from a project.
- More on Gold task
- Inter-annotator agreement (IAA)
- How often different people give the same label to the same item. Projects use it to measure how clear the guidelines are and how reliable each worker is.
- More on Inter-annotator agreement
- Consensus labeling
- Giving one item to several people and keeping the majority label.
- More on Consensus labeling
- Reviewer
- A senior contributor who checks other people’s submitted tasks and sends them back or approves them. Usually paid more than first-pass work.
- More on Reviewer
- Quality score
- A running grade of your work on a platform. It decides which projects you are offered, and on some platforms your pay tier.
- More on Quality score
- Hallucination
- A confident model statement that is false or made up, such as an invented citation. Spotting them is a core part of evaluation work.
- More on Hallucination
How the work is organised
- AI trainer
- A person paid to produce or judge data that trains AI models: writing answers, rating outputs, labeling data or testing for safety.
- More on AI trainerBrowse open roles
- Domain expert
- A trainer hired for specialist knowledge, such as a lawyer, doctor, engineer or PhD researcher. Expert roles pay well above general tasks.
- More on Domain expertHighest paying roles
- Project
- One client’s batch of work on a platform, with its own guidelines, pay rate and end date. Work comes and goes as projects open and close.
- More on Project
- Onboarding
- The unpaid or paid steps before real tasks: reading guidelines, passing a qualification test and sometimes a short trial project.
- More on Onboarding
- Assessment
- The entry test a platform gives applicants, often timed and in your field. Some platforms also run an AI video interview.
- More on AssessmentInterview prep
- EQ (empty queue)
- No tasks available on your project right now. It is common between batches and does not mean you were removed.
- More on EQ
- Task-based pay
- Pay per completed task rather than per hour. The effective hourly rate depends on how fast you work within the quality bar.
- More on Task-based pay
- Hourly pay
- Pay for time logged on a project, usually tracked by the platform. Most expert roles on this board quote an hourly rate.
- More on Hourly payPay stats
- Independent contractor
- How nearly all AI training platforms engage workers: you invoice or are paid per task, with no employment benefits, and you handle your own taxes.
- More on Independent contractor
- Referral link
- A sign-up link that credits whoever shared it if you are accepted. Many apply links on this board are referral links, which is how the board is funded.
- More on Referral linkAffiliate disclosure
Safety and policy
- AI alignment
- Making a model’s behaviour match what its makers and users actually intend. Human feedback is the main practical tool for it today.
- More on AI alignmentAI Safety & Evaluation roles
- Jailbreak
- A prompt crafted to get a model to ignore its safety rules. Red teamers write them so the model can be trained to resist.
- More on Jailbreak
- Content policy
- The rules for what a model may and may not say. Safety reviewers label answers against it.
- More on Content policy
- PII (personally identifiable information)
- Data that can identify a person, such as a name, email or ID number. Many tasks require removing or flagging it.
- More on PII
- Bias evaluation
- Checking whether a model treats groups of people unfairly, for example by testing the same prompt with different names or countries.
- More on Bias evaluation
Ready to look at roles?
Every listing shows its pay and who it is open to.