Lessons and practice quizzes are free with an account, and each examination includes two free attempts.
Foundation CAITR™ Foundation For everyone, no prerequisites. Take the modules in order: each one builds on the last.
- What AI training work is What SFT, RLHF, reward models and evaluation are, the main kinds of AI training task, and how a task moves from guidelines to review.
- Reading guidelines and onboarding How to work through a long project guideline, settle hard cases, ask about gaps instead of guessing, handle updates and get through onboarding.
- Annotation and labelling basics How to apply a label set consistently, draw a correct bounding box, know when segmentation is needed, and use unsure and skip options properly.
- Judging model answers How to rate a model's answer on instructions, accuracy, completeness, clarity, safety and format, and how to catch made-up facts.
- Writing for models How to write realistic, checkable prompts and good model answers, and how to rewrite a weak answer without adding new errors.
- Rubrics: using them and writing simple ones How to grade answers against a rubric consistently, criterion by criterion, and how to write short rubrics that others can apply the same way.
- Quality: scores, golden sets and review How platforms measure the quality of your work, how to act on reviewer feedback, and how to write justifications a reviewer can check.
- Working as a contractor: hours, pay and tax What contractor status means, how hours and pay are counted, how platforms pay, and where to find the official UK tax guidance.
- Getting hired safely The usual hiring steps on AI training platforms, the signs of a fake offer, how to check an address, and where to report a scam.
- Foundation final examination Timed: scenario questions and hands-on tasks, marked against criteria you see with each task.
Specialism Evaluation & RLHF For raters, evaluators and reviewers. 6 modules, about 7–10 hours. Counts as one of your 2 specialisms.
- Preference ranking and side-by-side comparison How to compare answers to the same prompt, choose consistently, use a strength scale with care and justify the choice in one checkable line.
- Rubric writing for evaluation How to write a rubric for an open-ended prompt: atomic, checkable criteria, must-avoid items and weights, tested on real answers and revised.
- Fact-checking and grounding in depth How to judge sources, check citations, quotes and figures, keep "unsupported" apart from "false", and record findings a reviewer can follow.
- Hallucinations and other failure patterns The common ways model answers fail, from invented references to sycophancy and needless refusals, and how to rate how much each one matters.
- Checking chain of thought How to follow a model's reasoning step by step, mark the first step that goes wrong and say why, and tell a lucky right answer from a sound one.
- Consistency, calibration and reviewing others How to keep your judgements steady over a long session, use golden-set results to recalibrate, and review other raters' work specifically and kindly.
- Evaluation & RLHF final examination Timed: scenario questions and hands-on tasks, marked against criteria you see with each task.
Specialism Expert Domains For professionals and graduates in a field. 5 modules, about 6–9 hours. Counts as one of your 2 specialisms.
- Writing hard prompts from your field How to write prompts from your field that are realistic, have a definite answer or marking standard, and are hard for current models.
- Reference answers and expert rubrics How to write a reference answer to professional standard and a rubric another expert would apply the same way, with criteria for safety-critical errors.
- Reasoning traces in your field How to write step-by-step working a model can learn from, and how to check a model's working and mark the first step that goes wrong.
- Professional boundaries Keeping employer, client and patient information out of tasks, checking your contract and professional rules, and not giving advice to people.
- Field pathway D1 to D3 applied to one field: the tasks each field sees, the mistakes models make in it, and a full prompt, reference answer and rubric set.
- Expert Domains final examination Timed: scenario questions and hands-on tasks, marked against criteria you see with each task.
Specialism Coding & Agents For software engineers and technical users. 5 modules, about 7–10 hours. Counts as one of your 2 specialisms.
- Evaluating code How to judge model-written code for correctness, edge cases, security, readability and fit, by running it, and how to compare two solutions.
- Coding prompts and reference solutions How to write coding prompts with a clear specification, tests that check it, and a clean reference solution, all of it your own work.
- Agentic tasks and trajectories What agentic tasks and trajectories are, how to grade a trajectory step by step, and how to write an agent task and a demonstration.
- RL environments and checks What an RL environment is, how to write tasks with automatic checks that only genuine solutions pass, and how to find and close reward hacks.
- Safe working practice Keeping secrets, personal data and client material safe on technical tasks, working in the sandbox provided, and logging time honestly.
- Coding & Agents final examination Timed: scenario questions and hands-on tasks, marked against criteria you see with each task.
Specialism Language & Audio For native speakers, linguists and transcribers. 5 modules, about 6–8 hours. Counts as one of your 2 specialisms.
- Transcription to a style guide How to transcribe speech exactly as a style guide asks: filler words, false starts, punctuation, numbers, speakers, overlaps, tags and timestamps.
- Recording speech data How to record utterances to a brief: a quiet room, the right distance and levels, checks before you submit, and consent and privacy for any voice.
- Labelling audio How to split audio by speaker, label sound events and noise, and rate audio quality and emotion consistently against a label set.
- Evaluating translation and localisation How to rate machine translation for accuracy, fluency and register, mark errors by type and severity, and fix a translation without changing its meaning.
- Rating model answers in your language How to judge a model's answer as a native speaker: natural phrasing, the right register, local conventions and content, and word-for-word translation.
- Language & Audio final examination Timed: scenario questions and hands-on tasks, marked against criteria you see with each task.
Specialism Safety & Red-teaming For writers, security and policy people, aged 18 or over. 5 modules, about 5–8 hours. Counts as one of your 2 specialisms.
- Harm categories and policies The harm categories most safety policies use, how a content policy is built, and how to apply one to borderline cases.
- Adversarial prompts and jailbreak patterns How adversarial prompts and jailbreaks work, explained at the level needed to test models, and how to write test prompts inside an authorised scope.
- Grading refusals and over-refusals How to judge whether a model was right to refuse, help in part or help fully, how to spot over-refusals, and how to rate a safe answer.
- Reporting findings How to write a red-team finding someone else can reproduce: setup, prompts, output, policy and severity, with harmful detail redacted.
- Wellbeing and limits Looking after yourself in safety work: the strain of distressing content, breaks and support, the right to step away, and the legal limits.
- Safety & Red-teaming final examination Timed: scenario questions and hands-on tasks, marked against criteria you see with each task.
Verify a CAITR™ credential · CAITR™ policies · Code of Conduct · © 2026 aitrainers.jobs.