hy270/ICT_Books

Rumi NBF/FBISE Grades 6–12 Page-Truth Corpus

Faithful, page-level capture of the National Book Foundation (NBF) / FBISE school textbooks for grades 6–12 used in the ICT (Islamabad) curriculum. Every printed page of 62 textbooks is paired with (a) a rendered page image and (b) a structured "page-truth" record that keeps the page's real reading-order layout — worked-out maths as drawn, tables in place, and Arabic text kept separate from Urdu.

Built to power retrieval / embeddings over textbook content at runtime and to reuse book figures as seed images for lesson-plan generation (HTML templates + image models), rather than re-generating every illustration from scratch.

⚠️ COPYRIGHTED SOURCE MATERIAL — RESTRICTED USE. Source textbooks are copyrighted (NBF/FBISE; some pages contain copyrighted literary excerpts). The repository was made public by the project owner on 2026-08-30, with that copyright status known, to let the wider team and community use it for education. The license remains restricted-educational: use for educational and research purposes per USAGE.md; do not redistribute the content as your own or commercially.

What's inside

• 62 books, grades 6–12, all subjects: mathematics, general science, physics, chemistry, biology, English, Urdu, Islamiat, history, geography, computer science, Pakistan Studies, zari taleem.
• ~11,150 pages, each with a rendered JPEG (200 DPI, ~1433 px wide) and a page-truth record.
• v1.1: 10,916 figure/table crops — every diagram, table, drawing and photo cut out of those pages as its own JPEG, with a description, layout coordinates and QA verdicts (data/figures.jsonl). See Figure & table crops below.

Layout

images/grade_<g>/<subject>/p<NNN>.jpg   # NNN = real printed page number
figures/<book_stem>/pg_<NNN>_f<k>.jpg   # v1.1: figure/table crops cut from the pages
data/pages.jsonl                         # one row per page (image path + page-truth + metadata)
data/pages-*.parquet                     # same rows, image bytes embedded (datasets-ready)
data/figures.jsonl                       # v1.1: one row per crop (description, bbox, QA verdicts)
metadata/books.jsonl                     # per-book metadata (offset, content_span, publisher, …)
images/metadata.jsonl                    # HF imagefolder convention (zero-code image loading)
DATASHEET.md                             # Datasheets-for-Datasets provenance
USAGE.md                                 # permitted-use / IP note

Row schema (one page)

field	meaning
id	grade_<g>_<subject>_p<NNN>
grade, subject, book_stem, publisher, medium, language	book/page provenance
printed_page_number, pdf_page_index	the real printed number + its index in the source PDF
image	rendered page JPEG (path in pages.jsonl; embedded bytes in parquet)
page_type, chapter, confidence	page classification + chapter context
text	flattened reading-order text — for embeddings / RAG
blocks	full structured JSON: headings / prose / lists / tables / worked-maths / dua / names, each language-tagged (ur, ar, en)
flags, needs_human_review	capture caveats; true where a fluent human should verify (Arabic diacritics, OCR pages)
source	vision-llm (captured by a vision model) · ocr-fallback (copyrighted literary pages, tesseract) · stub (genuinely blank/absent page)

Load it

from datasets import load_dataset
ds = load_dataset("hy270/rumi-nbf-fbise-6-12", split="train")   # parquet, image bytes embedded
# or images only, zero-code:
imgs = load_dataset("imagefolder", data_dir="images", split="train")

Figure & table crops (v1.1)

Every illustration, table, drawing and photo in the corpus, cut from its page as a standalone JPEG: 10,916 crops across all 62 books, under figures/<book_stem>/, one row each in data/figures.jsonl.

figs = load_dataset("hy270/ICT_Books", "figures", split="train")
good = figs.filter(lambda r: r["in_good_list"])   # the lesson-ready subset

Key row fields: book_stem, grade, subject, page, kind (illus/table), desc and text_in_image (from page-truth), bbox_norm (what was cut, normalised page coords), image (crop path), sha256, quality signals — flags, qa_status (the cutting drone's self-check), qa_final (an independent judge, where sampled), and in_good_list (the precomputed "safe to serve" verdict: self-check pass, no blocking flags, current cut config).

How they were cut: a vision model locates each page-truth block on a coordinate-ruler-burned page render; the box is then snapped to whitespace gutters, adjusted with caption-aware boundaries, and flagged (low_confidence, small_sparse, …). A drone self-checks every crop; independent judges on separate model accounts re-verify samples.

Measured quality (independent judges, fresh samples, final round n=141 across all books): right-figure 99%, complete 90%, description-match 92%; a self-check pass was judged correct 98% of the time. The self-check is deliberately over-cautious — most of its suspects are actually fine, so suspects are included with their flags rather than deleted; filter on in_good_list (or qa_status) for the conservative subset. Known characteristic: "clean" is the weak axis — many crops carry a sliver of neighbouring text or caption, because boundaries err toward never slicing the figure itself.

Intended use

Retrieval / embeddings over curriculum content, page-image reuse as seed images for lesson-plan rendering (v1.1 crops are exactly that seed set), layout/OCR research on multilingual (en/ur/ar) school textbooks.

Known limitations (read before training)

• needs_human_review pages (mostly Arabic/Qur'anic Islamiat pages + all ocr-fallback pages) are lower-fidelity — filter on the flag if fidelity matters.
• ocr-fallback pages are tesseract OCR (the vision model refused verbatim copyrighted text); expect OCR typos and flattened layout.
• stub rows have no image and empty text — genuinely blank/absent printed pages, or a few pages unobtainable from a defective source PDF (noted per-book in metadata/books.jsonl).
• Page-truth is model-captured, not hand-verified end to end. Treat as a strong draft.

See DATASHEET.md for full provenance, collection method, and preprocessing.