guangzhaoli avatar

wandb-weave

Query and analyze W&B experiment data and Weave LLM traces using Python scripts. Use when working wi

提供方 guangzhaoli|开源

wandb-skill

A Claude Code skill for Weights & Biases and Weave.

一个用于 Weights & Biases 和 Weave 的 Claude Code 技能。

Installation / 安装

git clone https://github.com/guangzhaoli/wandb-skill.git
cp -r wandb-skill ~/.claude/skills/wandb-weave

Prerequisites / 前置条件

pip install wandb weave
export WANDB_API_KEY="your-api-key"

How It Works / 使用方式

This skill automatically triggers when you ask Claude about:

当你向 Claude 询问以下内容时,技能会自动触发:

Trigger触发词
Query W&B experiments查询 W&B 实验
Analyze Weave traces分析 Weave traces
List projects列出项目
Create reports创建报告

Claude will automatically select and run the appropriate script.

Claude 会自动选择并运行合适的脚本。

Scripts / 脚本

ScriptPurpose用途
query_runs.pyQuery experiment runs with filtering and sorting查询实验 runs,支持过滤和排序
query_traces.pyAnalyze LLM traces by status, model, latency分析 LLM traces,按状态/模型/延迟过滤
list_projects.pyList all accessible entities and projects列出所有可访问的 entities 和 projects
create_report.pyGenerate shareable W&B reports生成可分享的 W&B 报告

Example Prompts / 示例提问

English:

  • "Query the last 10 finished runs in my-team/my-project"
  • "How many error traces are there?"
  • "Find LLM calls with latency over 5 seconds"
  • "Create a weekly training summary report"

中文:

  • "帮我查询 my-team/my-project 里最近完成的 10 个 runs"
  • "分析一下 traces 里有多少错误"
  • "找出延迟超过 5 秒的 LLM 调用"
  • "创建一个本周训练总结报告"

License

MIT