怎么读模型卡与系统卡How to Read Model Cards and System Cards

先分清对象：模型卡面向使用者（用途、限制、评测口径、训练数据说明），系统卡或部署安全页面向风险关注者（测了哪些风险、结果如何、什么条件下继续推进）。

看它测了什么、没测什么：只有能力评测、没有失效场景说明的文档，撑不起高风险用途；反过来只讲风险、不给可核对的口径也不行。

版本与日期一起记：同一型号在不同批次的卡片上结果可能不同，引用时把版本号与发布日期写进你的记录。

重点读「不适合做什么」那一栏：厂商写明的失败场景与拒答行为比宣传语可用；没有这一栏的卡片要打折看。

对照自己的用途找缺口：你要跑的场景如果不在它的评测覆盖里，这个缺口要自己补测，不能靠外推补齐。

举个例子（虚构需求）：输入是团队想用某个模型做医疗问答初稿 → 处理是把它的模型卡与系统卡并排读，列出「有没有医疗相关评测」「已知拒答与幻觉说明写在哪一栏」→ 结果判断是若两类文档都没覆盖该场景，结论写「卡片不足以支持此用途」→ 下一步是把该场景写进自己的评测样本，小范围试跑并留记录，再决定是否接入。

Tell the two apart by audience: model cards face users (intended use, limitations, evaluation scope, training-data notes); system cards and deployment safety pages face risk readers (which risks were measured, with what result, under what conditions the rollout continues).

Check what was measured and what was not: a document with only capability benchmarks and no failure-mode discussion cannot support a high-risk use; one that only discusses risk without checkable numbers is no better.

Record version and date together: results for the same model name can differ across card releases, so put the version and publication date in your notes when you cite it.

Read the "not suitable for" section closely: failure cases and refusal behaviour the publisher writes down are more usable than marketing lines; a card without that section deserves discount.

Map gaps onto your own use: if your scenario is outside its evaluation coverage, that gap is yours to test - extrapolation does not fill it.

Worked example (fictional need): input is a team wanting a model to draft medical Q&A answers -> processing reads the model card and system card side by side and lists "is there any medical evaluation" and "which section states known refusals and hallucination behaviour" -> the judgement is that if neither document covers the scenario, the conclusion reads "cards insufficient to support this use" -> next step: add the scenario to your own evaluation samples, run a small trial with records, then decide whether to integrate.