christian-bick/edugraph-bench

EduGraph Bench

Third-party learning tasks for developing and reviewing EduGraph's educational annotations. The initial seed contains 27 tasks from 3 documents and 2 publishers: 14 addition and 9 comparison tasks from Abby the Pup, and 4 Grade 1 story problems from first-edition Illustrative Mathematics material. These sources are separate from the project's synthetic material.

This release is a development seed, not an adjudicated benchmark or test set. Its labels are AI proposals: annotation_status is proposed, annotation_method is assistant_proposed, and labels is null. No human has adjudicated these rows. Do not report benchmark scores against proposed_labels as gold labels.

Contents and use

Each example represents one task, with an image crop, task and document identifiers, source attribution, and crop provenance. A task may have several related prompts when they form one inseparable activity. Retained context must include the instructions needed to interpret the task.

The initial annotation baseline is ontology 0.30.0 and annotation policy 0.1.0. This is an explicit baseline, not a claim that 0.30.0 is the newest ontology. Later ontology migrations must be separate, versioned changes.

from datasets import load_dataset

dataset = load_dataset("christian-bick/edugraph-bench", split="development")
example = dataset[0]
image = example["image"]
proposals = example["proposed_labels"]  # AI suggestions, not gold
assert example["labels"] is None

Images and ImageFolder metadata live under data/development/. The metadata retains sample_id, task_id, document_id, family_id, and source_id; source and license URLs; attribution; original-file and image hashes; page and crop coordinates; annotation versions; and exposure status. file_name associates each metadata row with its image. Consult the repository's schema for exact field types.

Sources and rights

Source	Admitted material	Upstream rights
Abby the Pup	Individually checked worksheets covered by its public-domain dedication	CC0 1.0; retain source credits and check exceptions
Illustrative Mathematics, first edition	Grade 1 Unit 2 Lesson 1 student worksheet, 2021	CC BY 4.0; separately credited images are CC0

The IM admission rule is edition-specific: IM's terms, section 7.1, identify the first edition published during 2019–2021 under CC BY 4.0. Current IM v.360 material is outside this seed's permissive-license policy. Abby's terms also require checking whether an exception is stated. A site's general license alone does not clear every embedded third-party asset.

Original assets retain their upstream licenses. Our original documentation, annotations, and collection rights are CC BY 4.0 only to the extent held by us and compatible with source obligations. Repository code is MIT. See LICENSE.md for scope and source policy for admission requirements. Neither a metadata label nor this dataset's title replaces item-level source terms.

Full credits are in ATTRIBUTIONS.md. The IM PDF's two Pixabay images were published before January 9, 2019 and are covered by Pixabay's CC0 provision, as recorded in asset-level provenance.

Reproduce the seed

python -m pip install -r requirements.txt
python scripts/fetch_sources.py
python scripts/build_seed.py --output temp/rebuilt/development
python scripts/validate.py
python -m unittest discover -s tests

The originals are included with SHA-256 pins; fetching verifies present files and downloads only missing originals. Crop recipes live in curation/tasks.json. Rendering is pinned to the dependency versions in requirements.txt, at 144 dpi. Coordinates are PDF points from the top-left, ordered [x0, y0, x1, y1]. Each crop combines the original instruction region (where needed), task region, and source footer with white padding. Rebuilding is local and makes no model calls. Use a new output directory to preserve any later human review work.

The ontology vocabulary in sources/ontology.json was exported from the published edugraph-py==0.30.0 package and includes 680 eligible local identifiers. The proposal sets are deliberately incomplete starting points for review, not exhaustively labeled examples.

Review and limitations

Use this seed to inspect the data format, develop annotation guidance, and organize human review. The annotation policy defines the path to gold labels. python scripts/validate.py checks the draft format; python scripts/validate.py --require-gold must reject this unreviewed seed.

Public source material may have appeared in model pretraining. Examples already used in the EduGraph demo are development material, and exposure status must be recorded without claiming that other examples are unseen. Group related pages, worksheet templates, and activity families before assigning any future splits. A separate source document does not by itself guarantee independence.

The seed is small and concentrated on early learning, selected publishers, English, and printable layouts. It does not establish coverage of all grades, subjects, teaching styles, languages, or accessibility needs. Human reviewers must check crop completeness and label correctness before promotion to a gold release.

Evaluation is an intended use. We request that benchmark users disclose training overlap and avoid training on any future designated evaluation split; this is a reproducibility request, not an additional restriction on upstream licensed rights.

Contributing

See CONTRIBUTING.md. Keep proposals, human decisions, and source evidence distinct. Candidate collection does not authorize publication until the source and release gates pass.