怎么读模型卡How to Read a Model Card
微型指南 · 约 6 分钟Guide · about 6 min · 引用核验 2026-11-03（仅引用链接与日期） · refs checked 2026-11-03 (reference links and dates only)
模型卡是厂商给模型的说明书。这份指南讲读它时该看哪几栏、哪些话要打折。A model card is the manufacturer's spec sheet for a model. This guide covers which sections to read and which claims to discount.
先看用途与限制栏：官方写明「适合什么、不适合什么」，比宣传语可靠。
看评测栏时先问「在哪些基准上测的」，没有基准的分数等于没有分数。
看训练数据栏：数据构成决定它可能带哪些偏见。
看许可栏：能不能商用、要不要署名、衍生模型要不要同许可。
看已知局限栏：厂商主动写的失败场景，比「效果拔群」更值得信。
对比多家时只比同口径：同一基准、同一设置，否则数字没有可比性。
Start with intended-use and limitations: what the maker says it is and is not good for beats marketing copy.
When reading benchmarks, first ask which ones were tested; a score without a benchmark is no score.
Read the training-data section: the data mix determines which biases it may carry.
Read the license: commercial use, attribution duties, and whether derivatives must share the same license.
Trust the known-limitations section: failure modes the maker admits are worth more than "state-of-the-art".
Compare vendors only on equal terms: same benchmark, same settings, or the numbers are not comparable.
例子与判断Example and interpretation
以下是教学情境，不是本站运行或实测记录。A teaching scenario, not a measured or executed site test.
读卡时做三栏笔记：已披露、有条件、未披露。只报告英文基准时，将中文效果放在未披露栏；有商用条款时记下版本与限制，不根据“开放”标签直接判断。Use three columns: disclosed, conditional and undisclosed. English-only benchmarks leave Chinese performance undisclosed. Record commercial terms and versions rather than inferring permission from an “open” label.
怎么向 AI 提问How to Ask AI · 怎么判断 AI 新闻是不是炒作How to Tell if AI News Is Hype · 什么时候不该用 RAGWhen Not to Use RAG · 它说得笃定，所以可信Confident Tone Means Correct