Skip to content
Back to skills

Llm Eval Gate

ASecurity

LLM 评测门禁:评测集、评分器、pass@k/pass^k 与回归判断

  • 39 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added November 12, 2026
ai-agents

Security analysis

A100/100

Scanned November 12, 2026

npx -y skills add oschina/mothx --skill llm-eval-gate --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Llm Eval Gate?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Llm Eval Gate
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/oschina-llm-eval-gate/badge)](https://www.skillsdirectory.com/skills/oschina-llm-eval-gate)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: llm-eval-gate
description: LLM 评测门禁:评测集、评分器、pass@k/pass^k 与回归判断
---

# LLM 评测门禁

## 何时使用

在修改提示词、模型、工具、检索、RAG 或 Agent 行为前后使用。

## 检查清单

### 定义

- 先写能力/回归评测定义与成功标准,再改实现;
- 评测集覆盖成功、失败、边界与对抗输入;
- 测试集不泄漏到训练、提示或示例。

### 评分

- 优先确定性检查,其次模型评分,最后人工;
- 模型评分说明偏差、抽样与一致性;
- 人工评分保留 rubric 与双人/复核机制(如适用)。

### 指标

- 报告 pass@k 与 pass^k,区分能力提升与回归;
- 记录模型、提示、工具、参数与数据版本;
- 同时报告成本、延迟与失败类型,不只看平均值。

### 门禁

- 明确发布阈值与回滚条件;
- 单次演示不能代替评测;
- 未运行或样本不足时明确标注,不报告为通过。

## 输出格式

- 评测定义与数据集;
- 评分器与运行命令;
- 基线与对比结果;
- 门禁结论与已知偏差。

## 反模式

- 用测试集调参后在同一测试集报告效果;
- 只报平均分,不报失败分布;
- 把“看起来更好”当评测结论。

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

View all in ai-agents →

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.