recall-bench

仓库创建 2026年6月6日最近提交 21 天前SkillHot 收录 20 天前

▸ 精选理由

为 RECALL 项目提供端到端可复现的基准与评判流程。

这个 Skill 做什么

在本地运行 RECALL 基准（含评判变体），并由调用代理担任评审汇总结果。

在本地跑完整的 RECALL 基准测试（含带评分的变体），并把分数聚合汇报。被调用的代理会担任评审角色，运行确定性 harness、收集结果并生成简要结论。仅在用户明确要求跑 benchmark 或对比基线时使用，不用于平常记忆访问或测试套件。

▸ 展开 SKILL.md 英文原文

Use this skill when the user explicitly asks to run, judge, or compare a RECALL benchmark (e.g. "run a judged normal bench", "recall-bench light", "benchmark RECALL against the baseline"). It runs the deterministic bench harness, optionally acts as the LLM judge for emitted judge tasks, aggregates scores, and reports results. Never invoke for ordinary RECALL memory usage, plugin development, or test-suite runs.

研究检索基准测试RECALL评测工具Claude Code

Stars

Forks

仓库内 Skill

7 日增星

安装 / 使用

给你的 Agent 一句话（通用）

帮我安装这个 skill：https://raw.githubusercontent.com/0langa/RECALL/main/.claude/skills/recall-bench/SKILL.md

或 curl 直取 SKILL.md

curl -fsSL "https://raw.githubusercontent.com/0langa/RECALL/main/.claude/skills/recall-bench/SKILL.md"

SKILL.MD 节选查看完整文件 ↗

# RECALL Benchmark Runner (repo-local skill)

Run the RECALL benchmark end to end, including the judged variant, and report
a concise summary. You (the invoked agent) are the judge — that is by design:
the harness makes zero LLM/API calls, and the user invokes this skill from a
session whose model they consider cheap enough to score rubrics.

Work from the repository root (the directory containing `bench/`). All
commands are local; nothing here contacts a network or needs credentials.

## Arguments

Parse from the user's request; defaults in bold:

- mode: `light` | **`normal`** | `complete`
- judged: **yes** if the user says "judged"/"with judge"; otherwise no
- baseline: compare against `b

via SKILL·HOT · 数据来自 GitHub 公开信息 · 原文版权归作者所有