Use this guide to research and maintain model knowledge cutoffs, family names, and generation labels in model catalogs. It explains how to rely on authoritative sources, derive labels from model IDs, and handle single-model updates or large backfills without guessing.
Use for model knowledgeCutoff, family and generation metadata: onboarding, corrections, research and bulk backfills.
Run score Not tested
See how to installThis guide explains how to research and maintain three model card fields: knowledgeCutoff, family, and generation. knowledgeCutoff is the point up to which a model learned world knowledge; family identifies its lineage; and generation identifies its version within that family. It directs you to use official documentation, model cards, and other authoritative sources, and to set values only when those sources or naming rules support them. It rejects guesses based on release dates and weak aggregator sites. The guidance covers both adding one model and filling fields across many models, including exceptions for derived models and generation labels.
Use it when adding a model to a catalog or correcting existing cutoff, family, or generation data. It also helps when reviewing a large catalog spanning multiple providers, so decisions about evidence and leaving fields empty follow the same rules.
Set values only when official information or naming rules support them. Leave fields empty when information is not public. Large research backfills use repository scripts and verification commands.
This overview was written by AI (OpenAI Codex) from the public SKILL.md. Check the original below for the exact behavior.
AI checked the SKILL.md for operations that need care and for setup requirements. The score comes only from actually running the skill (hands-on test).
Not counted in the score. Rated by AI (OpenAI Codex) as of Oct 26, 2026. Treat it as a guide and read the original before installing.
Review the instructions the agent will receive before you install.
How to populate and maintain the three structured metadata fields on packages/model-bank/src/aiModels/*.ts model cards, at single-model scale (new model PR) or repo-wide scale (sweep across ~80 provider files / ~1900 entries).
| Field | Format | Meaning |
|---|---|---|
knowledgeCutoff | 'YYYY-MM' (or 'YYYY' if only the year is published) | World-knowledge cutoff. When a vendor distinguishes a "reliable knowledge cutoff" from the broader training-data cutoff (Anthropic does), always use the reliable one. |
family | lowercase slug (claude, gpt, o-series, qwen, deepseek, llama, glm, …) | Model lineage, finer than organization. Lets the UI group models and match the same model across aggregator providers. |
generation | family slug + version (claude-4.6, gpt-5.2, qwen3.5, llama-3.1) | Generation within the family. Only set when confidently derivable from the model line's naming. Rolling aliases (qwen-max, deepseek-chat, gemini-flash-latest) get family only. |
All three are optional. The cardinal rule: only fill what an authoritative source states or naming rules derive — never guess. An empty field is correct for vendors that publish nothing.
No DB migration is ever needed for these: builtin models are merged from model-bank at read time (repositories/aiInfra/index.ts spreads the whole card), so new card fields flow to the client automatically.
Accept only:
Reject:
2024-06 for four distinct base models; none of the cited Cohere pages said that, and the only cutoff Cohere actually publishes is Feb 2023 for the 08-2024 Command R/R+ refresh.releasedAt — a release date is not a cutoff.Variant inheritance: dated snapshots (-2024-08-06), speed/price tiers of the same checkpoint, quantizations (-fp8, -awq), context-length variants (-32k), ollama :NNb tags, and cloud-prefixed ids (anthropic./us./global. Bedrock ids) share their base model's cutoff. Distills do not inherit from teacher or base — use the distill's own published value or leave empty. Sizes within one generation can genuinely differ: Llama 3 8B is Mar 2023 while 70B is Dec 2023 (per Meta's own card) — don't "fix" that to one family-wide value.
Vendors that publish no cutoffs (leave empty, don't chase): Qwen, DeepSeek, GLM/Zhipu, ERNIE, Doubao, Hunyuan, SenseNova, Spark, MiniMax, StepFun, Yi (mostly), Moonshot.
Known per-vendor footguns:
2025-05, Sonnet 4.6 is 2025-08 — easy to swap. Claude 3.7 is 2024-10 (system card: trained through Nov 2024, knowledge cutoff end of Oct 2024). Cite system cards / the models overview, not the Help Center article (a living page that drops retired models — citation rot).2023-04 vs gpt-4-0125-preview 2023-12).Rule-based, no research needed: .agents/skills/model-bank-metadata/scripts/derive-family.ts holds the per-family regex rules. Traps already encoded there — keep them when extending:
claude-sonnet-4-20250514 is generation claude-4, not claude-4.2.llama-3-8b → llama-3 (not llama-3.8); gemma-7b-it is gemma-1 (not gemma-7).qwen2p5 = qwen2.5, llama-v3p1 = llama-3.1, ollama :NNb tags, Bedrock us./global./anthropic. prefixes.claude-X.0 normalizes to claude-X.claude-fable-5) don't match the opus/sonnet/haiku regex — they are the Mythos class — family: 'claude-mythos', generation: 'mythos-5' (set manually; the launch page calls Fable 5 "the generally available Mythos-class model").bun .agents/skills/model-bank-metadata/scripts/extract-model-ids.ts → unique normalized chat-model ids (normalization = last path segment, lowercased). Non-chat types (image/video/embedding/tts) have no knowledge cutoff — skip them.{id, cutoff, source} with the sourcing rules above baked into the prompt, plus one adversarial verify agent per chunk that re-fetches cited sources and refutes unsupported claims. The verify pass is load-bearing: it caught the Cohere aggregator copy-paste and the AWS launch-date conflation.sources map — e.g. drop everything sourced to aws.amazon.com).bun .agents/skills/model-bank-metadata/scripts/apply-cutoffs.ts <map.json> and bun .agents/skills/model-bank-metadata/scripts/apply-family.ts <map.json> (run from repo root). Both are idempotent codemods keyed on normalized id — aggregator providers get the same values automatically; entries that already have the field are skipped. They rely on the uniform prettier formatting of the data files (entries start { / end },, fields at 4-space indent).cd packages/model-bank && bunx vitest run src/aiModels/__tests__/index.test.ts && bunx tsc --noEmit.anthropic.ts for reference values).git grep -c knowledgeCutoff -- 'packages/model-bank/src/aiModels/*.ts' before vs after. A three-way stack of model PRs once silently dropped all 10 Anthropic cutoffs during conflict resolution.No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Use this skill to verify a delivery by exercising the real product through a suitable surface such as a CLI, web app, desktop app, or iOS Simulator. It helps create or find a verification plan, capture evidence, and publish an acceptance report.
Use this skill to add an AI model provider to LobeHub across runtime integration, model cards, settings, branding, documentation, and acceptance checks. It also covers provider documentation updates and distinguishes new providers from models added to an existing provider.
Use this skill to add a server-side environment variable that sets a default for a user setting. It walks through validation, server configuration and store merging, plus updating the example environment file and self-hosting documentation.
Provides guidance for registering hooks around an agent’s execution steps, tool calls, human approvals, context compression, and sub-agent calls. Hooks can observe events, supply local mock tool results, or use webhooks to allow or deny tool results under configured policies.
Build event-driven background workflows for agents by defining sources, interpreting them as signals, and running actions through policies. It also guides asynchronous handoffs, deduplication, scoped execution, completion receipts, and observability.
Use this skill to run end-to-end acceptance checks for bot channels in real desktop chat apps. It drives apps such as Discord, Slack, Telegram, WeChat, Lark, and QQ to send a message, wait for a response, and capture the result.