DreamsHunter/ncert-curriculum-sft

NCERT Curriculum SFT Dataset

Lineage

This dataset is the output of the EduData-Kit pipeline (Project 1 of a 6-project ML portfolio for the Pexcera AI Engineer role). The data flow is:

1. NCERT textbook PDFs (grades 6–12, science & mathematics) — sourced via ncert.overrides (explicit per-chapter URLs) and/or live scrape of the NCERT textbook index. Scanned PDFs are flagged and excluded (no OCR).
2. Heading-aware chunking with pdfplumber (font-size ratio 1.15) and a 1200-token cap; target 3000–5000 chunks.
3. Teacher-LLM Q&A distillation with Gemini 2.5 Flash (Google AI Studio paid tier, billed through GCP — $300 credits). 5 Q&A per chunk, Bloom-tagged. Paid-tier limits (1000 RPM / 1M RPD) let ~1400 chunks complete in ~2 minutes for ~$1.50. Uses Gemini's native JSON mode for structured output; the parser also defensively strips code fences and recovers from trailing-comma errors.
4. Off-the-shelf MCQ merge — SciQ, OpenBookQA, and MMLU-STEM (high-school biology/chemistry/physics/mathematics + elementary mathematics, conceptual physics, astronomy), filtered by a syllabus-aligned topic_whitelist.
5. Near-duplicate removal — paraphrase-multilingual-MiniLM-L12-v2 embeddings + FAISS greedy drop at cosine > 0.92.
6. Quality filter — min answer 20 chars, min question 8 chars, refuse-marker drop, per-chapter cap of 200.
7. Format + split — messages schema, stratified 95/5 on class, plus a frozen MCQ hold-out (≤300 rows) pushed to a separate eval repo.

Chapters covered

NCERT grades 6–12, science & mathematics. The exact chapter list is determined by ncert.overrides in config.yaml (the reliable path) plus whatever the live scrape discovers. See outputs/pdf_manifest.json for the authoritative list of chapters actually downloaded and parsed.

Off-the-shelf sources

source	dataset	split used	role
SciQ	sciq	train + validation + test	MCQ eval pool
OpenBookQA	allenai/openbookqa	train + validation + test	MCQ eval pool
MMLU	cais/mmlu	test (STEM subjects)	MCQ eval pool

License

• NCERT textbook text: public / Crown (Government of India) — used as source passages for teacher-LLM distillation, not redistributed verbatim.
• Generated Q&A + messages format: MIT (yours — produced by this pipeline).
• SciQ: CC BY-NC-SA 4.0.
• OpenBookQA: CC BY 4.0.
• MMLU: MIT.

Stats

Fill these in after the real run (python src/07_format_messages.py logs per-class counts; the EDA notebook plots the distributions). The table below is a template — replace the numbers with your actuals.

split	rows	notes
train	20714	95% stratified on class
test	1090	5% stratified on class (held out)
eval_mcq	170	frozen — never trained on (Project 5)

class	train	test
6	2110	111
7	2107	111
8	3393	179
9	2726	143
10	1838	97
11	3555	187
12	4985	262

subject	rows
science	14080
mathematics	7724

Sample rows

Replace with 5 actual rows from outputs/final_dataset/train.jsonl after the run. Example shape:

{"messages": [{"role": "system", "content": "You are a patient CBSE curriculum tutor. Explain concepts clearly using simple, grade-appropriate language, step-by-step. Always ground your answer in the curriculum concept the student asks about. Keep answers focused and pedagogical."}, {"role": "user", "content": "Why does an increase in the number of units being shared lead to a larger share for each child, assuming the number of children does not change?"}, {"role": "assistant", "content": "An increase in the number of units leads to a larger share for each child because the total amount available to be divided has grown. With more items to distribute among the same fixed number of children, each child receives a greater individual portion of the total units."}], "metadata": {"class": 6, "subject": "mathematics", "chapter": "fractions", "source": "ncert-distilled"}}

Known limitations

• NCERT scraping fragility. The pipeline's live scrape of the NCERT textbook index is best-effort; markup changes break it. Populate ncert.overrides for reliability.
• No OCR. Scanned PDFs are flagged (is_scanned=True) and skipped — no text is extracted from them in this build.
• Off-the-shelf MCQs are open-grade. SciQ/OBQA/MMLU items have no fixed grade; they enter the frozen eval hold-out, not the grade-stratified SFT split.
• Teacher-LLM generation. Gemini-distilled answers are grounded in the source passage by prompt design, but may occasionally contain phrasing artifacts. The quality filter (step 6) drops refuse-markers and short answers, but does not verify factual correctness against an oracle.
• Dedup is question-level. Two questions with the same meaning but different surface form may survive if their cosine is ≤ 0.92.

Citation

This dataset was built as part of the EduData-Kit portfolio project. If you use it, cite the EduData-Kit repository and the relevant source datasets (SciQ, OpenBookQA, MMLU). The LoRA / knowledge-distillation methodology referenced by the downstream projects (EduTutor-LoRA, Paper2Code-LoRA-Distill) is described in the portfolio's planning documents (02_EduTutor-LoRA.md, 04_Paper2Code-LoRA-Distill.md).