Skip to content

ML Model Card

A model card records an ML model's capabilities, limitations, training provenance, disaggregated evaluation results, and ethical considerations in a standardised, stakeholder-facing format. Model cards enable cross-team trust, satisfy emerging transparency obligations under the Foundation Model Transparency Index and the EU AI Act, and give Forensic Auditors a single canonical artifact to verify when a model is promoted through an FCC Critique. Produce this artifact during the Critique phase whenever a model is trained, fine-tuned, or re-evaluated against new data.

Template

Section 1: Model Overview

Instructions: Identify the model, its architecture, base (for fine-tuned models), owning team, license, and related Innovation / patent IDs. Keep the summary to 1-2 sentences.

Field Value
Model name [FILL]
Version [FILL]
Model type [classifier / regressor / generative / embedding / other]
Architecture [FILL]
Base model [FILL — if fine-tuned]
License / IP status [FILL]
Summary (1-2 sentences) [FILL]

Section 2: Intended Uses

Instructions: Distinguish primary uses from downstream uses (where the model feeds another system) and out-of-scope uses. Explicit out-of-scope statements are mandatory — silence is not consent.

  • Primary use cases: [FILL]
  • Downstream uses: [FILL]
  • Out-of-scope uses: [FILL]

Section 3: Training Details

Instructions: Link to the Dataset Card (OPEN-SCI-004b) for every training source. Declare hyperparameters, selection method, compute footprint, and CO₂ estimate (use the ML CO2 Impact Calculator or an equivalent). Sensitive-data handling must be explicit.

  • Training datasets (link to Dataset Card): [FILL]
  • Preprocessing: [FILL]
  • Hyperparameters + selection method: [FILL]
  • Hardware, training time, cost, CO₂: [FILL]

Section 4: Evaluation

Instructions: Report headline metrics and disaggregated metrics by relevant subgroup (demographic, domain, temporal). Include confidence intervals or error bars; bare point estimates are not sufficient for Critique sign-off.

  • Headline metrics: [FILL]
  • Testing data + split method: [FILL]
  • Disaggregated results: [FILL]
  • Statistical significance (CI, p-value, method): [FILL]

Section 5: Bias, Risks, Limitations

Instructions: Enumerate known biases, risks, and limitations in plain language. Recommendations give consumers the complementary practices that close the residual-risk loop.

  • Known biases: [FILL]
  • Risks: [FILL]
  • Limitations: [FILL]
  • Recommendations: [FILL]

Section 6: Provenance & Reproducibility

Instructions: Cross-link to the preregistration (OPEN-SCI-001), the reproducibility checklist (OPEN-SCI-003), the training code repository + commit hash, experiment tracker run ID, and the random seed(s).

  • Preregistration ID: [FILL]
  • Reproducibility checklist ID: [FILL]
  • Code repo + commit: [FILL]
  • Experiment tracker run ID: [FILL]
  • Seeds / environment: [FILL]

Adoption Checklist

  • All required sections completed
  • Artifact peer-reviewed by at least one R.I.S.C.E.A.R. peer
  • Stored in the project's designated docs location
  • Linked from README or equivalent index
  • Versioned + date-stamped alongside the model artifact it describes

References

  • PHOENIX v4.0.0 — docs/resources/templates/open-science/model-card.md
  • Mitchell, M. et al. (2019) — Model Cards for Model Reporting, FAT* '19
  • Hugging Face — Annotated Model Card Template
  • Bommasani, R. et al. (2023) — Foundation Model Transparency Index, Stanford CRFM
  • EU AI Act (Regulation 2024/1689) — Article 13, Annex IV

FCC integration

This template is referenced from the Forensic Auditor persona (src/fcc/data/personas/forensic_auditor.yaml) as part of the Critique-phase evidence set. Model cards are the primary artifact the auditor inspects when verifying that a model promotion meets the transparency and risk-documentation obligations of the EU AI Act mapping under src/fcc/data/compliance/eu_ai_act_requirements.yaml. FCC also auto-generates 173 baseline model cards; this hand-authored template is used when the generated card needs human-readable supplementation. See also src/fcc/data/governance/open_science_gates.yaml.