AI Quiz Generator — turn notes into a multiple-choice quiz

The generator turns free text into multiple-choice questions with a two-track pipeline. The deterministic track is always available — it runs in your browser, needs no network, and produces the same quiz every time — while an optional neural track adds exam-style question stems when the server has a Hugging Face key configured. Nothing you paste is stored.

1. Segment. Your text is whitespace-normalised and split into sentences with a Unicode-safe splitter that respects . ? ! and skips common abbreviations (Dr., e.g., etc.). Empty and duplicate sentences are dropped.
2. Score key terms. In each sentence every token is scored: a number scores highest, then a capitalised proper noun, then the longest content word (stop-words never win). The top-scoring token becomes the blank; the same rule builds the pool of distractors from the other sentences. Because the rule is fixed, the same sentence always yields the same question — and a second, independent scorer cross-checks every key term on every run.
3. Build the MCQ.The blanked term is the correct answer. Distractors are drawn first from other sentences' key terms of the same kind (numbers for a number, names for a name), then from numeric or textual padding so the option count is always met. The correct option is placed at slot (index + 1) mod options — reproducible, and spread evenly so the answer is not always “A”.
4. True/False & difficulty. For true/false, alternate sentences are kept true; the rest are corrupted into a plausible false variant (a number is nudged, a copula is negated, or a key term is swapped). Each question is tagged Easy, Medium, or Hard by mapping its leading verb to a Bloom's Taxonomy cognitive level.
5. Neural stems (optional). In Multiple-choice mode, when a key is set, the cleaned text is sent once to google/flan-t5-base through the Hugging Face Inference API — the same server-side pattern the site's summarizer and question-answering tools use. Parsed stems are repaired so each keeps exactly one correct answer, merged ahead of the deterministic ones, and if the model is unavailable the deterministic quiz stands on its own.
6. Take, score & export. Take-quiz scoring is pure array logic — one point per stem whose selected option equals the answer index. Exports are pure string serialisations built in your browser: GIFT (::title:: stem { =right ~wrong #### feedback }), Aiken, RFC-4180 CSV, JSON, and a printable student copy or answer key. Every special character is escaped so Moodle parses the file.

Up to 8,000 characters can be processed at once. Generated questions are a study aid, not a source of truth — always check them against your material, especially for exam revision, and edit or delete anything that is off.

• Moodle Docs — GIFT format (question, answers, and #### feedback syntax)
• Moodle Docs — Aiken format (stem, lettered options, ANSWER line)
• Vanderbilt CFT — Bloom's Taxonomy (cognitive levels used for difficulty tags)
• google/flan-t5-base — the instruction-tuned model used for neural question stems
• RFC 4180 — the CSV field-quoting spec the CSV export follows

The export-format and taxonomy references above were last cross-checked on 2026-07-10. The deterministic fill-in-the-blank and true/false worked examples on this page are reproducible from the tool exactly.