md. I’ll also keep carrying the specific repository + license notice on every walkthrough page so each doc makes its licensing context explicit.
Dataset Card Template¶
Metadata¶
- Dataset name:
<UKB/ABCD/etc> - Source / accession:
<URL or data agreement> - License / DUA:
<e.g., UKB Research Analysis Tool terms> - Primary contact:
<PI / analyst> - Related YAML:
kb/datasets/<id>.yaml
Access & Storage¶
- Paths to raw data (e.g.,
/secure/ukb/raw), processed derivatives, and sync locations. - Encryption / PHI considerations; who has access.
Hugging Face / Shared Mirrors¶
- Dataset repo + subset names, record counts, checksum or snapshot tags.
- Notes on sync cadence between mirrors and secured copies.
Cohort Summary¶
| Split | N subjects | Notes |
|---|---|---|
| Train | <N> |
<filters> |
| Val | <N> |
|
| Test | <N> |
Add stratification notes (sex balance, age range, sites).
Modality-Specific Schema¶
- Column lists per modality (genes, ROIs, clinical tables) with links back to
docs/data/schemas.md. - Required covariates + units.
Overlap & Pairing Logic¶
- How subjects overlap with other modalities (e.g., gene × sMRI intersection).
- Inclusion/exclusion rules, QC flags (motion, missing genes, etc.).
Genomics / Base-Pair Stats¶
- Total base pairs, mean/median sequence length, tokenizer context alignment details.
Preprocessing Pipelines¶
- Imaging: software versions, atlases, smoothing, censoring thresholds.
- Genetics: variant calling, phasing, annotation, gene list definition.
- Clinical: diagnosis coding, questionnaire scoring.
Available Features & Covariates¶
- Feature tables (e.g., FreeSurfer ROIs, FC z-maps, gene embeddings).
- Covariates recommended for residualization (age, sex, site, motion, PCs, SES).
- Missingness patterns and imputation strategy if required.
QC & Harmonization¶
- Thresholds (FD < 0.3 mm, min TR length).
- Site/scanner harmonization (e.g., ComBat) or reasons for not applying.
- Logs / reports location.
Notes & Risks¶
- Any data use restrictions (no redistribution, publication review).
- Confounds or biases to monitor (ancestry imbalance, site skew).
- Planned updates or reprocessing tasks.
References¶
- Link to cohort publications / documentation.
- Internal tickets or notebooks for preprocessing runs.
Linked Walkthroughs & External Repos¶
- Walkthrough pages that describe preprocessing.
external_repos/*scripts or configs that produced the released tensors (commit/tag + entrypoint).