Handwritten exam answer-sheet pages with verbatim machine-OCR transcriptions and preserved spatial annotations, exported from GradeLab production data (anonymized institutional exam program).

Structure

data/pages/train|validation|test/train-000NNN.tar — WebDataset shards. Each sample: {sample_id}.jpg (full page) + {sample_id}.json
data/blanks/train/blanks-train-000NNN.tar — genuine blank pages (negative class for blank-page detection / OCR robustness)
data/regions/figures/regions-figures-000NNN.tar — figure crops: {crop_id}.jpg + {crop_id}.json linked to the parent page sample
index/pages.parquet, index/regions.parquet — metadata indexes
dataset_manifest.json — counts, checksums, provenance

Sample JSON fields

schema_version, sample_id, document_id, page_number, page_type (content|blank), text (complete page-level OCR transcription — verbatim), ocr_confidence, word_count, image (dims/sha256/bytes), answer_regions (per-question answer_box, normalized 0–1000 [yMin,xMin,yMax,xMax] + answer_box_px), figures (description, bbox_px, normalized_bbox, linked crop_id), cancelled_regions (bbox, cancelled text, cancellation_type, classification, confidence, reason), label_source, label_quality, notes.

Coordinate conventions (two pixel orders exist — each field declares its own)

Normalized: [yMin, xMin, yMax, xMax] on a 0–1000 scale (0,0 = top-left).
answer_regions[].answer_box, cancelled_regions[].bbox are normalized (originals); their *_px companions are pixel y-first [yMin,xMin,yMax,xMax].
figures[].normalized_bbox is normalized y-first; figures[].bbox_px is the production-stored crop box in pixel x-first [x1, y1, x2, y2] (same frame used to cut the stored crop images).
Every annotation carries an explicit *_coordinate_system field — never assume ordering across annotation types.

Label quality

All transcriptions are unverified machine OCR output (label_quality: unverified_machine_ocr) produced by google/gemini-3.5-flash via OpenRouter. They are NOT human-validated ground truth. Question IDs / answer_regions are provided for optional layout and region supervision only — do not use them as training targets.

Splits

Deterministic, grouped by document (pdf hash): train 315103, validation 3008, test 3193 pages. No document appears in two splits. Includes 22674 blank pages as a separate negative class.

Privacy

Student identifiers were removed; pages with detected identifiers are quarantined and excluded.
Pseudonymous IDs only; the raw source mapping is kept outside this repo.
Private dataset — do not redistribute.

Loading

import webdataset as wds
url = "pipe:curl -sL https://huggingface.co/datasets/<repo>/resolve/main/data/pages/train/train-000000.tar"
ds = wds.WebDataset(url).decode()
for sample in ds:
    img, meta = sample["jpg"], sample["json"]

Downloads last month: 233

Total file size:

107 GB