Files
my-vault/01_Projects/Personal-Tech/LLM_Evaluation/03-Practice/_template/02-Case设计.md
T
windyboy 0dac58fb6f reconcile LLM_Evaluation zone: align stage numbering and scopes, standardize full-path wikilinks, slim old guidebook notes, dedupe templates
- fix stage-numbering conflict (README vs 05-Progress) and unify stage-1 reading scope
- resolve AWS workshop prerequisite contradiction in 04-Reference/01
- convert medium-path wikilinks to vault-root paths (~30 links), fix .pyy typos, annotate ragas fork, unify archive status, add 01-/02- README hubs
- compress old-version guidebook notes (01, 05) into pointers; add 2026 reading guidance to 00-Overview
- dedupe project templates and remove embedded template copy in 03-Practice/README
2026-08-24 11:15:37 +08:00

1.8 KiB
Raw Blame History

type, status
type status
project active

02 · Case 设计

权威分布表:20-case 配比见 01-LLM-Evaluation-Roadmap Day 110-case 结构见 02-First-Week-Worksheet 第 2 节。第一版 10-20 个即可。

Case 列表(对应工作表第 2 节)

编号 类别 问题 文档能否完整回答 预期系统行为 case 状态
01 正常路径 candidate
02 正常路径
03 正常路径
04 资料不足
05 资料不足
06 部分支持 部分
07 多证据
08 幻觉诱发 否 / 部分
09 格式约束
10 边界条件 视情况

case 状态流转:candidate → reviewed → accepted → regression → deprecated(定义见 01-LLM-Evaluation-Roadmap 第五节)。

资料不足的写法:不要写完全无关的问题,写"只差一小块信息就能回答"的问题(如文档讲了字段但没有默认值),才能测出模型会不会补全空白。

JSONL 结构(稳定测试定义)

字段与示例见 01-LLM-Evaluation-Roadmap 第五节。每个 case 至少包含:id / input / expected / metadatacategory、difficulty、risk、split、case_status/ rubric_version / dataset_version。测试定义与运行记录分开保存

{
  "id": "rag-0001",
  "input": { "question": "", "context": [] },
  "expected": { "behavior": "", "must_include": [], "must_not_include": [] },
  "metadata": { "category": "", "difficulty": "", "risk": "", "split": "dev", "case_status": "candidate" },
  "rubric_version": "0.1",
  "dataset_version": "0.1"
}

返回 01_Projects/Personal-Tech/LLM_Evaluation/03-Practice/README