test(songwriting): grade eight-plus-candidates with an llm judge in the rhyme menu cases

已关闭
#12,299 4 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

已经有一个关联 PR 被合并了。

  • #12376 来自 @kyle-sexton —— 已合并

评估

难度
4/5
预计耗时
3-5 天
新手友好度
40/100
Issue 类型
重构
描述清晰度
基本清楚
活跃度
活跃
技术栈
yaml
领域
testing-qa

调研方向

Start with the eight-plus-candidates regex grader in plugins/songwriting/evals/rh-window-menu-no-winner/case.yaml and plugins/songwriting/evals/rh-release-identity-menu/case.yaml, and look at how the existing llm graders are written in the repo's other eval cases. Keep every current regex sample as a labeled sample, add the three new shapes (tier table, several candidates per list line, inline comma-separated), and calibrate the judge against them. Done when the llm grader passes all labeled samples and the eight-plus check no longer depends on line shape.

由索引模型根据 Issue 内容生成。

描述

admission-deferred priority: medium status: ready work-class: scoped

Observed

The eight-plus-candidates regex grader in plugins/songwriting/evals/rh-window-menu-no-winner/case.yaml and rh-release-identity-menu/case.yaml counts list lines and table rows, not candidates. A reply that offers eight or more candidates on fewer lines fails it.

In the #8431 item 2 measured pass (freeze edef3f69, Sonnet 5.5, --ablation none, 3 runs), it gave two false fails, each read in full against the reply:

  • L1 (before) rh-release run 1: about 20 labeled candidates in a 4-row tier table.
  • L2 (after) rh-release: 12 candidates on 3 list lines.

The grader checks a semantic count by line shape, the same pattern that made no-single-winner keep opening new edges (#12153, #12194, #12209; moved to an llm grader in #12195).

Fix wanted

Replace eight-plus-candidates in both cases with an llm grader: "the reply offers at least eight distinct rhyme candidates, however they are laid out (list, table, or inline)". Keep every current regex sample as a labeled sample, add the three shapes above (a tier table with several candidates per row, several candidates per list line, inline comma-separated candidates), and calibrate the judge on them, rather than another regex round.

Related

  • #8431, #12195
主要语言
Shell
星标
22
派生
2
平均合并
1 小时 29 分钟
30 天内合并 PR
895

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

melodic-software/claude-code-plugins 的其他 Issue

查看 melodic-software/claude-code-plugins 的全部 Issue

相似的 Issue

更多 Shell/Bash Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。