VixeroAI/spm-synthetic-questions

Dataset Card: SPM Synthetic Questions

Dataset Summary

14,135 synthetic exam-style questions for Malaysia's SPM (Sijil Pelajaran Malaysia) curriculum, Form 5, covering 10 subjects. Every item is LLM-generated and aligned to the KSSM curriculum and the SPM examination format. Each item belongs to one of three categories:

• —hots — Higher Order Thinking Skills (KBAT) questions
• —lazim — soalan lazim (commonly-asked question styles)
• —perangkap — soalan perangkap (trap questions testing common mistakes)

Categories are near-balanced (⅓ each) and languages are near-balanced between Bahasa Melayu and English (½ each). Reference answers are provided for every question. This is a training corpus (SFT), not an evaluation benchmark. A larger top-up release is planned.

Intended Use

• —Supervised fine-tuning and post-training of language models for Malaysian secondary-school exam reasoning.
• —Research on bilingual (ms/en) educational question generation.
• —Tutoring and exam-preparation tooling as a source of practice questions.

Answers are model-generated reference answers, not official mark schemes. Treat them as supervision signal, not ground truth.

Dataset Structure

Single split: train (14,135 rows).

Example row:

{
  "id": "c3b-000001",
  "subject": "Additional Mathematics",
  "form": 5,
  "category": "hots",
  "language": "bm",
  "question": "Diberi persamaan kuadratik ...",
  "answer": "(a) Hasil tambah punca = ..."
}

Subject distribution

Category counts: hots 4,929 / lazim 4,459 / perangkap 4,747. Language counts: bm 7,166 / en 6,969.

Dataset Creation

All items are synthetically generated by a large language model, conditioned on the KSSM Form 5 curriculum and the SPM examination format (question styles, paper structure, answering conventions). Generation ran with per-item validation and automatic retry; items failing validation were regenerated. No external exam content is included or reproduced in this dataset.

Post-generation quality filtering removed rows containing promotional content, generator artifacts, and malformed or degenerate items, followed by exact and question-level deduplication. Rows were then shuffled and assigned stable identifiers.

Limitations

• —Answers are not official. Reference answers are model-generated and were not verified against any official scheme. Error rate is nonzero, especially on multi-step numerical working.
• —Subject imbalance. The largest subject is ~9.3× the smallest (Bahasa Melayu vs Geography). Per-subject weighting may be needed during training.
• —Form 5 only. Form 4 content is not included in this release.
• —Residual near-duplication. Exact duplicates were removed, but common question templates (e.g. standard solve-for-x phrasings) recur across rows by nature of the exam format.
• —Synthetic provenance. All items are LLM-generated; stylistic fingerprints of the generator model may be present.

License

Apache 2.0.

Citation

@dataset{vixero_spm_synthetic_questions_2026,
  title        = {SPM Synthetic Questions},
  author       = {{Vixero Technology Enterprise}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/VixeroAI/spm-synthetic-questions}},
  license      = {Apache-2.0}
}

Contact

Vixero Technology Enterprise — Kuching, Sarawak, Malaysia.