Skip to content

Card Template

Every model card produced by the plugin conforms to a single template: model-cards/templates/MODEL_CARD.md. This page reproduces the template structure for reference — it is what the validation checkpoint in /model-card create checks against, and what hand-authored cards should match.

Frontmatter

---
model_name: <provider>/<model-name>
provider: <Anthropic | OpenAI | Google | Meta | Mistral | xAI | Cohere | Alibaba | other>
model_version: <as-stated-by-provider>
last_researched: YYYY-MM-DD
card_version: 0.1.0
researcher: model-card-researcher (claude-opus-4-7[1m])
sources:
  - tier: 1
    url: <provider docs URL>
    fetched: YYYY-MM-DD
  - tier: 2
    url: <huggingface URL or "n/a">
    fetched: YYYY-MM-DD
  - tier: 3
    url: <arxiv URL or "n/a">
    fetched: YYYY-MM-DD
  - tier: 4
    url: <web URL or "n/a">
    fetched: YYYY-MM-DD
---

Required keys

Key Description
model_name The model's canonical identifier as published by the provider.
provider The provider company name.
model_version Version as stated by the provider. If the provider does not disclose a separate version, use the input model name.
last_researched Date the agent (or hand-author) last fetched sources, in YYYY-MM-DD.
card_version The card schema version. Always 0.1.0 for this plugin's v0.1.0.
researcher Identifier for who or what produced the card. Agent uses model-card-researcher (claude-opus-4-7[1m]).
sources One entry per tier consulted. Tiers not consulted use url: "n/a". The validation checkpoint requires at least the tiers actually cited in the body.

Body — the ten sections

The body must contain all ten sections in canonical order, with the exact heading text shown below. Sparse sections are filled with "Not publicly available" rather than omitted — the validation checkpoint enforces presence, not depth.

# Model Card: {provider}/{model-name}

## 1. Model Details

What this model is, who released it, when, what kind of model it is. Cite
provider-stated identity, release date, model family, parameter count if
disclosed.

## 2. Intended Use

Primary uses, primary users, out-of-scope uses (per the provider's published
statements).

## 3. Factors

Relevant subgroups, environmental factors, instrumentation factors. Often
sparse for proprietary API-served models — fill with "Not publicly available"
where the provider has not disclosed.

## 4. Metrics

Performance measures the provider reports. Cite each metric with its source
tier; prefer tier-1 (provider docs) and tier-3 (release paper).

## 5. Evaluation Data

Datasets used to evaluate (per the provider's published statements). Often
"Not publicly available" for proprietary models — say so explicitly.

## 6. Training Data

Datasets used to train (per the provider's published statements). Frequently
"Not publicly available" for proprietary models — say so explicitly.

## 7. Quantitative Analyses

Disaggregated performance, where provider has published this. Cite metrics
table from release paper or provider docs; record the date of the analysis.

## 8. Ethical Considerations

Risks, mitigations, known failure modes — preferring the release paper as
primary source (tier 3); supplemented by provider statements (tier 1) and
independent evaluations (tier 4) where relevant.

## 9. Caveats and Recommendations

What this card cannot tell you (sections marked "Not publicly available" and
why). Recommendations for use given the model's documented strengths and
limitations. Conflicts between tier-1 and tier-4 sources are recorded here.

## 10. Operational Details

Best-effort structured fields; missing values are written as "Not publicly
available" rather than omitted:

- **Pricing** — input/output cost per million tokens; pricing tier; as-of
  date. Cite the pricing page (tier 1).
- **Context window** — maximum input + output tokens.
- **Knowledge cutoff** — provider-stated training cutoff date.
- **Supported APIs/SDKs** — provider's API name(s); language SDKs; OpenAI
  Chat Completions compatibility if applicable.
- **Latency tier** — provider's stated latency tier or category, if any.
- **Tool / structured-output support** — function calling, JSON mode, tool
  use, parallel tool calls.
- **Function calling** — supported, parallel-supported, schema format.
- **Fine-tuning availability** — yes/no/limited; cite the fine-tuning docs
  page if available.
- **Rate-limit notes** — provider-published rate limit tiers or quotas.

Section 10 — field-level "Not publicly available"

Section 10 (Operational Details) is special: every field listed in the template appears in every card, even if the value is unknown. Field-level "Not publicly available" entries replace silent omission.

- **Pricing** — Not publicly available [T1.4]
- **Context window** — 1M tokens [T1.2]
- **Knowledge cutoff** — January 2026 [T1.1]
- **Fine-tuning availability** — Not publicly available

This makes downstream comparison reliable — the future /model-card compare (issue #234) can diff cards field-by-field without having to special-case missing fields.

Validation checkpoint

The validation step in /model-card create checks four things before writing:

  1. YAML frontmatter parses; required keys are present.
  2. All 10 numbered section headings are present in canonical order (## 1. Model Details through ## 10. Operational Details).
  3. Every per-claim citation matches \[T[1-4]\.\d+\] and resolves via the frontmatter sources block.
  4. Section 10 fields use field-level "Not publicly available" rather than silent omission.

Deviations are fixed in place — the agent is not re-dispatched. Citation gaps (e.g. a [T2.1] reference with no source 2.1 in frontmatter) are surfaced to the user before write.