AI Question Paper Generator from Syllabus

Setting a good question paper is slow, skilled work: the paper must cover the whole syllabus proportionally, mix recall with analysis-level questions, and avoid repeating last year's paper. Faculty do it by hand, and question banks go stale. This project automates the drafting stage — not the final judgment. Given syllabus text, a T5 question-generation model fine-tuned on SQuAD-derived (context, question) pairs drafts candidate questions from each syllabus section; a concept extractor ensures every key topic is covered; and a Bloom's-taxonomy classifier labels each question (remember, understand, apply, analyze) so the difficulty mix is deliberate. A blueprint table shows exactly how marks distribute across topics and cognitive levels, and the teacher edits the draft into the final paper. The demo generates a complete 70-mark paper from a sample operating-systems syllabus in seconds.

1. Syllabus text is segmented into topics; a TF-IDF + noun-phrase extractor identifies the key concepts per topic.
2. For each concept, the surrounding syllabus passage and the concept are fed to the fine-tuned T5 question-generation model, which drafts candidate questions.
3. A Bloom's-level classifier (DistilBERT fine-tuned on labeled question data) labels each generated question's cognitive level.
4. A selection stage picks questions to satisfy the teacher's difficulty mix and total marks while maximizing concept coverage and minimizing near-duplicates.
5. The blueprint table is computed (marks per topic per Bloom's level) and the paper renders in printable format with sections and instructions.
6. The demo evaluates generation quality with BLEU/METEOR against SQuAD references and shows a human-rated sample in the report.

Tech stack:

• Python 3.11, Hugging Face Transformers
• T5-base fine-tuned for question generation
• SQuAD 2.0 (Stanford, 100k+ QA pairs)
• DistilBERT (Bloom's-level classifier)
• scikit-learn (TF-IDF concept extraction, metrics)
• Flask + HTML/CSS/JS (demo)
• PyTorch (fine-tuning)

Dataset & model details

• Dataset: SQuAD 2.0 (Rajpurkar et al., Stanford, 2018) — 100,000+ question–answer pairs on Wikipedia articles, restructured into (context, question) pairs for question-generation training. Source: rajpurkar.github.io/SQuAD-explorer.
• Task: Question generation — input = syllabus passage + target concept, output = natural exam question; plus Bloom's-level classification of generated questions.
• Model: T5-base (220M parameters) fine-tuned on SQuAD-derived QG pairs; DistilBERT classifier for Bloom's taxonomy levels (4 classes).
• Metrics: BLEU/METEOR for generation quality against SQuAD references; Bloom's classifier accuracy on labeled questions — design targets reported after training.

• [T5 question generation] T5 fine-tuned on SQuAD-derived pairs drafts natural questions grounded in the provided syllabus text.
• [Concept coverage] Key concepts are extracted per syllabus section and every concept gets question coverage — no silent topic gaps.
• [Bloom's taxonomy labeling] Each question is classified into Remember / Understand / Apply / Analyze so the difficulty mix is measurable.
• [Marks blueprint] A blueprint table shows marks distributed across topics × cognitive levels, the standard exam-design artifact.
• [Difficulty mix control] The teacher sets the target mix (e.g. 30% recall, 40% understanding, 30% application/analysis) and the generator fills it.
• [Printable paper format] The draft renders as a clean, formatted exam paper with sections, marks, and general instructions.
• [Live web demo] Paste syllabus text or use the preloaded OS syllabus; the full paper generates instantly with its blueprint.

• Fine-tuned T5 question-generation model
• Concept extraction + Bloom's classification + blueprint pipeline source
• Web demo with preloaded operating-systems syllabus
• QG evaluation notebook (BLEU/METEOR, human-rated samples)
• Difficulty-mix configuration presets
• Project report PDF (background, question generation, Bloom's taxonomy, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (T5, SQuAD, BLEU, Bloom's taxonomy)

• Generated questions are drafts — a teacher must review, edit and approve the final paper; the project never claims exam-ready output without human review.
• BLEU/METEOR are design targets until training executes; the report records achieved figures plus human ratings.
• The model generates from the given syllabus text only; it cannot invent questions about topics not present in the input.
• Numerical/problem-style questions (derivations, circuit analysis) are weaker than theory questions — documented in the limitations.
• Demo syllabus is operating systems; other subjects work the same way but need their text pasted in.