Practice tasks for AI training work
Short tasks modeled on AI training work. Each one is new, and you see what you missed as soon as you check.
Compare an answer with its source, one sentence at a time. Some tasks have no problem at all.
Check the answer against the source
A model was asked to answer from this source only. Label each sentence of its answer. Pick Contradicts when the source says something different, and Not in source when the source doesn’t say.
Question: What is happening with the light rail extension?
- The extension is not open yet.
- It should open in 2030.
- The line will end at the university.
- It will have 6 new stations.
- A new sales tax will pay for it.
Your weak spots
Source check, saved on this device
Today’s task
Everyone gets the same source check today, at Standard level. A new one starts at midnight UTC.
Each task matches a kind of project work
Source check
Grounded answering and fact-checking projects ask you to judge each claim in a model answer against a document. The hard part is telling a claim the source contradicts from a claim the source never makes. F5: Ground Truth Research Method teaches the method.
Side-by-side
Comparing two responses to one request is the task most generalist projects start with. A losing response usually misses part of the request, gets a fact wrong or takes the wrong tone, and the rating needs the reason as well as the pick. F6: Advanced Instruction Following teaches the method.
Summary check
Summarization projects ask whether a model kept every point the source makes and changed none of them. A dropped point is the hard case, because nothing on the page is wrong, and you only find it by reading the source again. F5: Ground Truth Research Method teaches the method.
Style guide edit
Writing-quality projects give you a style guide and ask you to find each line that breaks it. The rules change from project to project, so the guide gets read first and the text is checked against it, line by line. CW1: Creative Writing & Style Guide Adherence teaches the method.
Math check
Math projects ask you to check each step of a model’s working. An early slip carries through every later step, so the first wrong step is the one to report. STEM1: STEM & LaTeX (Math Solving) teaches the method.
JSON check
Annotation projects often export labels as JSON. One trailing comma breaks the whole file, and a value of the wrong type passes the parser but fails the schema. F7: Structuring & Formatting Data for AI teaches the method.
For multiple-choice questions and the timed practice assessment, use the practice hub.
Common questions
Are the practice tasks free?
Yes. They need no account, and your results stay on your own device.
Do practice tasks count toward the AI Trainer Certification?
No. They train the skills the certification exam tests, and the modules linked under each task teach the method. The exam is a separate timed test.
How close are these to the tasks on AI training platforms?
They follow the shapes that recur across platforms: compare two responses, check an answer against a source, catch the schema error. Live projects add their own guidelines and tools on top, so treat these as the drill, and the guidelines of each project as the test. The guide toAI training task examples walks through seven real ones.
Why do some tasks have nothing wrong with them?
On Standard and Hard, about one task in five is clean. Real queues include correct answers, and marking a correct answer wrong costs you accuracy the same way a missed error does.
Can I send a task to someone else?
Yes. Each task has a "Copy link to this task" action, and the link replays the same task. The task of the day is the same for everyone and changes at midnight UTC.