Skip to content
PEBEducational Services

How the data is built

Methodology

One pipeline, applied across four modalities. Every stage below states what it removes, and the per-dataset sections show how each modality applies it.

Pipeline overview

Every record, in every dataset, passes through these six stages in order.

  1. Acquisition

    Sources material from publishers, repositories, CAD archives, and chart generation.

  2. Rights verification

    Confirms a licensing or ownership basis before intake.

    Removes material without a confirmed rights basis.

  3. Normalization

    Converts source formats into each modality's delivery schema.

    Removes files that fail to parse or normalize cleanly.

  4. Deduplication

    Compares hashes and near-duplicates per modality.

    Removes redundant and near-identical records.

  5. Contamination screening

    Checks against named public benchmark test splits, per modality.

    Removes records that overlap a benchmark test or validation split.

  6. Human audit sampling

    Manually reviews a sampled subset before a release is tagged.

    Removes records that fail manual review.

Text

Textbook Dataset

How the shared pipeline applies to this dataset, and the benchmarks it is screened against.

Quality pipeline

  1. Ingestion and text extraction

    Extracts full text per page and reconciles the result against the printed page count.

    Files that fail extraction or show a page-count mismatch.

  2. ISBN validation

    Verifies the ISBN-13 checksum and confirms a catalog match.

    Invalid or non-existent ISBNs.

  3. Duplicate detection

    Compares ISBN and normalized title across the catalog.

    Exact duplicate records.

  4. Language identification

    Classifies language at the page level and reconciles against the declared language field.

    Records with unresolved language mismatches.

  5. Metadata completeness check

    Confirms title, category, and page count are present and non-null.

    Records with incomplete required fields.

Contamination screening

Benchmarks checked
MMLU, GSM8K, MATH, ARC-Challenge, HellaSwag
Method
13-gram overlap screening against benchmark test and validation splits, consistent with decontamination practice used across large-scale pretraining corpora.

Code

Coding Dataset

How the shared pipeline applies to this dataset, and the benchmarks it is screened against.

Quality pipeline

  1. Repository and problem intake

    Pulls candidate repositories via the GitHub API; authors DSA problem-solution pairs to a fixed rubric.

    Inactive, archived, or low-signal repositories.

  2. License audit

    Confirms an OSI-approved or permissive license before inclusion.

    Repositories with restrictive or unclear licensing.

  3. Execution verification

    Runs every DSA solution against its test cases per language.

    Non-compiling or failing submissions.

  4. Duplicate detection

    Compares repository full name and normalized problem title.

    Duplicate entries.

  5. Static security scan

    Scans dependencies and code patterns against known vulnerability signatures.

    Files flagged for known vulnerable patterns.

Contamination screening

Benchmarks checked
HumanEval, MBPP, SWE-bench, LiveCodeBench
Method
Function-signature and docstring matching against benchmark problem sets, with manual review of high-similarity flags.

3D geometry

3D CAD Dataset

How the shared pipeline applies to this dataset, and the benchmarks it is screened against.

Quality pipeline

  1. Geometry validation

    Confirms the file parses cleanly in its declared format and is not corrupt or empty.

    Files that fail to parse.

  2. Domain classification

    Assigns primary domain and category via the source catalog taxonomy.

    Records the taxonomy cannot place.

  3. Deduplication

    Compares geometry and mesh hashes for near-duplicate assets.

    Redundant near-duplicate assets.

  4. Topology repair check

    Flags non-manifold or self-intersecting geometry.

    Assets with unrepaired topology defects.

  5. License and rights audit

    Confirms a redistribution basis before inclusion.

    Assets without a confirmed rights basis.

Contamination screening

Benchmarks checked
ABC Dataset test split, ShapeNet-Part, Fusion 360 Gallery
Method
Geometric hash and near-duplicate mesh comparison against published benchmark test splits.

Image + structured data

Infographics Dataset

How the shared pipeline applies to this dataset, and the benchmarks it is screened against.

Quality pipeline

  1. Data-to-image generation

    Renders each chart or diagram from its structured source, or ingests a cleared redrawn editorial image.

    Source data that fails to render cleanly.

  2. Pair verification

    Confirms the structured data and rendered image describe the same values.

    Mismatched image/data pairs.

  3. Category classification

    Assigns main category and subcategory via the source taxonomy.

    Records the taxonomy cannot place.

  4. Visual deduplication

    Compares perceptual hashes across the catalog.

    Near-identical renders.

  5. Legibility check

    Flags low-resolution or overlapping-label images.

    Images that fail the legibility threshold.

Contamination screening

Benchmarks checked
ChartQA, PlotQA, InfographicVQA, DocVQA
Method
Perceptual hash and structured-data overlap comparison against published benchmark test splits.