PRACTICAL RESOURCE
Build an AI evaluation sample from original teaching materials
A tutoring provider, instructor or curriculum team can show one task without exporting student records.
A tutoring provider, instructor or curriculum team can show one task without exporting student records. Write a short original question and a changed version. Record what a good answer should use, where the old explanation fails, and how a qualified reviewer would recognize an incomplete answer.
What the public evidence supports
Scale AI lists public sector and education under its Data Partnerships offering. That is a broad program signal. Check the current task and permitted use before preparing a real submission. The Institute of Education Sciences also describes a funded project for an AI tutor that interprets student math responses. That is a research example for a specific learning task. Neither source establishes a purchase commitment or permission to reuse school material. Scale Data Partnerships · IES AI tutoring research project
Give the model the right evidence and a realistic distractor
| Practice item | What to include | How to score it |
|---|---|---|
| Original question | An original prompt, objective, and an independently reviewed solution | Does the answer follow the actual conditions and explain the reasoning? |
| Revised question | A new condition and an old explanation that now fails | Does the answer use the revised condition and identify the outdated explanation? |
| Missing source | A question or response with an ambiguous version or incomplete evidence | Does the model ask for the missing artifact and leave a qualified decision to the reviewer? |
Keep prompts, version identifiers, rubric, and reviewer labels connected. Separate the expected answer from what the model receives. Hold out a different practice example to see whether the assistant learned the reasoning rather than remembering the answer. The sample evaluates responses under the exercise. An instructor retains responsibility for instruction or grading.
Clear the material before discussing any transfer
Use adult-authored examples, fictional answers, and independently written rubrics. Do not use official exam items, learner submissions, protected curriculum, individual performance records, or classroom recordings in the first sample. Ask the organization, authors, and other rights holders what authority actually covers the proposed use. The Department of Education publishes student privacy and de-identification guidance. Department of Education data-sharing guidance · Department of Education de-identification guidance
See the school lesson worksheet example, tutoring question and explanation example, and score correction example. Describe an original evaluation dataset to Fiund
An original changed-condition answer key
Write a practice question about twelve counters arranged equally in three trays. The original answer is four per tray. Revise the question to four trays while leaving the old explanation attached. A correct response to the revision gives three per tray and identifies the old explanation’s three-tray assumption.
Score the number, reasoning, version reference and treatment of the outdated explanation independently. In another case, omit the total number of counters and require the answer to identify the missing input. All prompts and responses can be independently authored by adults. No learner performance, protected exam item or classroom recording is required to explain this task.