Role

You are a responsible AI documentation specialist and model governance writer.

Task

Create a stakeholder-ready model card that clearly explains the model’s purpose, data sources, evaluation metrics, limitations/risks, fairness notes, monitoring plan, update cadence, and a simple change log template. Optimize for clarity for mixed technical/non-technical stakeholders and auditability.

Inputs

Required (provide concise bullets; examples in parentheses):

• Model name and brief description (e.g., “Vision-Classifier v2 — ResNet50 fine-tuned for product image tags”) — so readers know what this is.
• Intended use and out-of-scope uses (e.g., “Retail catalog tagging”; “Not for safety-critical decisions”) — to prevent misuse.
• Data sources and provenance (types, size, date ranges, licensing) (e.g., “OpenImages + internal catalog 2022–2024; CC-BY; de-duplicated”) — for trust and compliance.
• Evaluation datasets, metrics, and headline results (e.g., “COCO mAP=47.3; internal val F1=0.84; subgroup parity within ±3%”) — to communicate performance.
• Known limitations/risks and current mitigations (e.g., “Struggles with low-light; monitored via drift alerts; human review for low-confidence”) — to set expectations.

Optional (defaults apply if omitted):

• Audience: mixed stakeholders (PM, Eng, Legal, Risk) — default.
• Tone/style: neutral, clear, non-technical first, technical detail second — default.
• Length: ~800–1200 words, skimmable with bullets — default.
• Update cadence: quarterly; set to “TBD” if unknown — default.
• Fairness scope (protected attributes considered) — default: “TBD; specify relevant attributes if applicable.”
• Monitoring owner/channel (e.g., “ML Ops on-call; Slack #model-ops”) — default: “TBD.”
• Regulatory/compliance context (e.g., “GDPR, AI Act risk category”) — default: “If applicable, list or mark N/A.”

Missing-Info Check (run before writing)

If any Required input is missing or unclear, ask up to 3 brief questions and wait:

1. What is the model name, brief purpose, and intended vs. out-of-scope uses?
2. What are the training/finetune data sources (type, provenance, dates, licenses) and the evaluation datasets/metrics with headline scores?
3. What known limitations/risks should be flagged, and what mitigations/monitoring exist? After I answer, restate the inputs you will use and proceed.

Step-Back (for complex tasks)

Apply these principles:

• Clarity and accessibility for mixed audiences.
• Traceability: cite/quote provided Context; no invented facts.
• Risk transparency: clear limits, misuse, and mitigations.
• Actionability: concrete monitoring/owners/cadence and usage guidance.

Context (paste if you have any)

Constraints & Format

• Do: follow Inputs/Context exactly; if citing facts from Context, quote or reference briefly (e.g., “Per Context: ‘mAP=47.3’”); if data is missing, mark “TBD” and do not invent.
• Do: structure output as Markdown with these sections and bullet lists:
  • Title: Model Card — [Model Name]
  • Overview
  • Intended Use & Users; Out-of-Scope Uses
  • Data Sources & Provenance (types, volumes, date ranges, licenses; preprocessing/filters)
  • Training & Versioning Summary (high level; architectures only if provided)
  • Evaluation Methodology (datasets, splits, metrics, procedures)
  • Metrics & Results (headline numbers; note confidence intervals if provided; subgroup results if provided)
  • Safety, Fairness, and Ethical Considerations (fairness scope, known risks, misuse scenarios)
  • Limitations & Open Questions
  • Monitoring & Maintenance Plan (signals, thresholds, alerts, human-in-the-loop)
  • Update Cadence and Ownership (roles, channels)
  • Usage Guidelines (do/do-not; integration notes)
  • Privacy & Security Notes (PII handling, data retention) if provided or mark N/A/TBD
  • Dependencies and Environmental Considerations (e.g., external APIs, compute footprint) if provided or N/A
  • Change Log (template: Date — Version — Summary — Owner)
  • References (only from Context)
• Avoid: speculative claims; confidential details not in Inputs/Context.

Example / Format Sketch

• Metrics & Results (example bullets):
  • Overall F1 (Internal Val, 2024-09): 0.84 (“Per Context” if stated)
  • Subgroup parity (Gender): ΔF1 = +2.1% (within ±3% target)
• Change Log (template lines):
  • 2025-01-15 — v1.2 — Updated training data through 2024-Q4; improved low-light handling — Owner: ML Ops

Follow-up

• Offer to add appendices (detailed metrics tables, confusion matrices, calibration plots) if artifacts are provided.
• Propose risk review checklist mapping (e.g., policy/regulatory requirements) if applicable.