53-Model Card Template v1.0 | Chapter 5 — Training Data & Lineage

Home ／ Docs-Technical WhitePaper ／ 53-Model Card Template v1.0

Chapter 5 — Training Data & Lineage

I. Purpose & Scope

• Specify sources, schema & splits, lineage, and license compliance for training data so that train/val/test are fully consistent with the rest of the model card in units/dimensions, coverage mode, versioning, and freshness.
• For path quantities (arrival time/phase), the text must explicitly show gamma(ell) and d ell; the data side records delta_form ∈ {general, factored}; parenthesized unified forms are required; publication requires p_dim = 1.0.

II. Inputs & Dependencies

• Dataset Card: align with Ch. 3/4/6/7/8/10/11/12 (provenance.yaml, schema.json/contract.yaml, split.yaml/split_manifest.json, validate_report.json, report_manifest.yaml).
• Parameter Card: align freshness.policy and cov_group; any calibration/constant used for training must be registered and traceable.
• Error Budget Card: unify coverage mode (k/alpha/quantile) and covariance config (Σ, kernels & params).
• Pipeline Card: inbound contracts and stage interfaces; training/validation/export paths & release layout.
• Citations & versions: “volume + version + anchor (P/S/M/I)”, anchor coverage ≥ 90%, no external links/aliases.

III. Sources & Licenses

• Types: instrument/system/simulator/external; in data_refs.yaml record source_id, producer, license, site, operator.
• License/limits: record license and redistribution terms; if usage/region/audience limits exist, state them in the model card front matter and API docs.
• Time lock & freshness: clock_state="locked", |ts_start − calib.timestamp| ≤ τ_calib; isolate expired samples or tag [Restricted].

IV. Schema & Splits Alignment

• Fields/units/dimensions: training-read fields must match Dataset Card Ch. 4; missingness as null/omitted with reasons in quality.flags.
• Leakage prevention: strictly follow split.yaml time/entity isolation and RNG seed; forbid cross-split entity/window sharing.
• Path consistency: len(gamma_ell)=len(d_ell)=len(n_eff)≥2; step Δell ≤ ( c_ref / f_s ) / max(n_eff); align phase in the reference window before metrics.

V. Sampling & Cleaning

• Sampling: random/stratified/hard-first recorded with RNG and quotas in data_refs.yaml (strata by batch/device/region/quality.flags).
• Cleaning rules: missing/outlier/denoise/normalize in preprocess_spec.yaml; any distribution-altering rule is logged in Bias (Ch. 8 Dataset Card) and evaluated by slices.
• Augmentation/synthesis: record provenance, params, and ratios; count separately in split_manifest.json.

VI. Lineage & Traceability

• Lineage DAG: raw → calibrated → derived → annotated → split → train_batch, nodes/edges labeled with version & checksum; cycles forbidden.
• Event audit: acquisition/clean/split/augment/sample events to audit.jsonl (time, operator, input hashes, change note, signature).
• Reproducibility: train_config.yaml pins data paths/versions/snapshots; provide a minimal replay script reproduce.sh in the appendix.

VII. Normative Path Forms

• Arrival (two equivalent):
  T_arr = ( 1 / c_ref ) * ( ∫ n_eff d ell )
  T_arr = ( ∫ ( n_eff / c_ref ) d ell )
• Phase:
  Phi = ( 2π / λ_ref ) * ( ∫ n_eff d ell )

In text, explicitly show gamma(ell) and d ell; on data side record delta_form; training/eval path/phase conventions must match the Dataset Card.

VIII. Gate Mapping

• G1 Schema completeness: training-read fields match contracts.
• G2 Citation compliance: anchor coverage ≥ 90%.
• G3 Path conventions: gamma/measure/delta_form complete; step & alignment compliant.
• G4 Dimensional closure: end-to-end I70-dim_check passed, p_dim = 1.0.
• G5 Freshness: clock_state="locked", τ_calib valid.
• G6 Coverage consistency: training/evaluation use the same k/alpha/quantile.
• G7 Covariance consistency: Σ PD and aligned with the Error Budget.
• G8 Uniqueness & acyclicity: unique record_id/checksum, Lineage DAG acyclic.
• Any core-gate failure triggers S1–S5 (dimension/freshness/path/covariance/citation), blocking training & release; tag [Restricted] when necessary.

IX. Machine-Readable Artifacts
A. data_refs.yaml

version: "1.0.0"

datasets:

- id: "ds-core"

see:

- "Dataset Card v1.0:Ch.3"

- "Dataset Card v1.0:Ch.4"

- "Dataset Card v1.0:Ch.6"

manifest: "DS_EXPORT/manifests/report_manifest.yaml"

splits: "DS_EXPORT/splits/split_manifest.json"

license: "CC-BY-4.0"

checksum: "sha256:..."

sampling:

seed: 20250924

strategy: { stratified: ["device","region","quality.flags"] }

preprocess_spec: "configs/preprocess_spec.yaml"

B. preprocess_spec.yaml

version: "1.0.0"

missing: { numeric: "null", route_to: "quality.flags" }

normalize: { mean: "μ_train", std: "σ_train" }

path_align: { require: true, delta_form: "general", enforce_delta_ell: true }

filters:

- name: "window_guard"

rule: "drop if ts ∉ [ts_start, ts_end]"

audits: { write_to: "reports/audit.jsonl" }

C. lineage_graph.json (excerpt)

{
  "nodes": [
    { "id": "RAW-telemetry", "version": "1.0.0", "checksum": "sha256:..." },
    { "id": "CAL-telemetry", "version": "1.0.1", "checksum": "sha256:..." },
    { "id": "DER-train", "version": "1.0.0", "checksum": "sha256:..." }
  ],
  "edges": [
    { "from": "RAW-telemetry", "to": "CAL-telemetry", "type": "calibrate" },
    { "from": "CAL-telemetry", "to": "DER-train", "type": "derive" }
  ]
}

X. Anti-Patterns & Fixes

• Anti: T_arr = ∫ n_eff / c_ref d ell (missing parentheses) → Fix: parenthesize to the unified form.
• Anti: training uses only gamma(ell) without d ell/delta_form → Fix: complete and equalize with n_eff; reject if non-compliant.
• Anti: split leakage (entity across splits) → Fix: recut with group_by(entity) and update split_manifest.json.
• Anti: expired samples or unlocked clocks → Fix: filter per freshness.policy or isolate; tag [Restricted] if needed.
• Anti: inconsistent coverage between training and evaluation → Fix: unify a single coverage.mode and declare in manifests.

XI. Cross-References

• Dataset Card: Ch. 3 (Provenance), Ch. 4 (Schema), Ch. 6 (Splits/Versioning), Ch. 7 (QC Gates), Ch. 8 (UQ & Cov), Ch. 11 (Bench/Score).
• Error Budget Card: Ch. 5/6/8 (covariance & coverage), Ch. 9 (threshold mapping).
• Parameter Card: Ch. 4/6/8/9 (units/freshness/covariance groups/interfaces).
• Pipeline Card: Ch. 3/4/6/12 (graph/contract/stages/release).

XII. Checklist

• data_refs.yaml / preprocess_spec.yaml / lineage_graph.json ready and consistent with Dataset Card manifests.
• For path quantities, explicit gamma/measure/delta_form; len(path) ≥ 2, Δell compliant; phase aligned in reference window.
• I70-dim_check passed, p_dim = 1.0; coverage k/alpha/quantile unified with Error Budget.
• Leakage prevention and license/usage verification complete; audit.jsonl fully recorded.
• /validate passed G1–G8; non-compliances marked [Restricted]; citation anchor coverage ≥ 90%.