🤖 Agentic Data Science Competition

🤖 智能体驱动的数据科学竞赛

Turn AI Agents into Kaggle Teammates
让 AI 智能体成为你的 Kaggle 队友
_{Not just tools — autonomous collaborators that research, debug, iterate, and win.
不仅仅是工具 —— 能够自主研究、调试、迭代并获胜的合作者。}

Quick Start • Key Insights • Case Studies • Install

---

Distilled from real Kaggle competition experience 提炼自 真实 Kaggle 竞赛实战经验

Including: RL Game AI • Audio Classification • LLM Reasoning • Multiple debugging journeys 包括：强化学习游戏 AI • 音频分类 • LLM 推理 • 多次调试实践

---

🎯 What This Skill Does / 这个 Skill 能做什么

Transform your Kaggle workflow from manual iteration to autonomous agent-driven competition:

将你的 Kaggle 工作流从手动迭代转变为智能体驱动竞赛：

Before 之前	After 之后
Manual notebook analysis 手动分析 notebook	Agent pulls top solutions with dependencies 智能体拉取顶级方案及依赖
Guess why submission failed 猜测提交失败原因	Agent diagnoses 400 errors, zip format issues 智能体诊断错误
Wait and refresh for scores 刷新等待分数	Cronjob monitors kernel, auto-submits Cronjob 监控，自动提交
Try random improvements 随机尝试改进	Spec-driven development → delegate → verify 规范驱动开发

---

⚡ Quick Start / 快速开始

# Install the skill / 安装 skill
mkdir -p ~/.hermes/skills/data-science/agentic-kaggle/
curl -sL https://raw.githubusercontent.com/FrankS-IntelLab/agentic-kaggle-skill/main/SKILL.md \
  -o ~/.hermes/skills/data-science/agentic-kaggle/SKILL.md

Then in your AI agent / 然后在你的 AI 智能体中：

> Use the agentic kaggle skill to help me replicate this top notebook
> Why is my submission returning 400 error?
> Set up auto-monitoring for my kernel

---

💡 Key Insights / 核心洞察

2️⃣ Competition Types Matter / 竞赛类型很重要

Type 类型	Submit 提交内容	Detection 检测方法
Answer 答案	CSV predictions CSV 预测	Most competitions 大多数竞赛
Model 模型	LoRA/checkpoints LoRA/检查点	Top notebooks train models 顶级 notebook 训练模型

3️⃣ Kernel Mode Trap / Kernel 模式陷阱

Mode 模式	Test Set 测试集	Result 结果
kaggle kernels push	❌ Hidden 隐藏	Invalid submission 无效提交
"Save & Run All"	✅ Mounted 挂载	Valid submission 有效提交

4️⃣ Top Replication Workflow / 顶级方案复制流程

# Pull WITH metadata (-m is critical!) / 带 metadata 拉取（-m 很关键！）
kaggle kernels pull user/top-notebook -p ./solution/ -m

# Edit only id/title, KEEP all dependencies / 只修改 id/title，保留所有依赖
kaggle kernels push -p ./solution/

---

📊 Case Studies / 案例研究

RL Strategy Game Competition / RL 策略游戏竞赛

Lesson 教训	Details 详情
Feature completeness 功能完整性	Top agents: 3,000+ lines → LB 1200+ 顶级智能体：3000+ 行
Simplified agents 简化版	~120 lines, 4/12 features → LB 500-600 简化版：~120 行
Time budget 时间预算	Strict turn limits — profile after each change 严格回合限制

👉 Full Case Study / 完整案例

Audio Classification Competition / 音频分类竞赛

Lesson 教训	Details 详情
Hybrid ensemble 混合集成	Temporal model + SED ensemble = Top scores 时序模型 + SED 集成
Prior limitations 先验局限	Location priors don't help when all samples are similar 样本相似时先验无效
Silent failures 静默失败	Log exceptions during feature extraction 记录异常

👉 Full Case Study / 完整案例

---

🛠️ Troubleshooting Cheat Sheet / 故障排除速查表

Problem 问题	Solution 解决方案
400 Bad Request	Try .zip format (only zip the CSV!) 尝试 .zip 格式
FileNotFoundError	Check /kaggle/input/competitions/<name>/ 检查路径
Training IDs in submission 提交包含训练 ID	Use sample_submission.csv for fallback 使用 sample_submission.csv
Score dropped 分数下降	Wait 4h for stabilization 等待 4 小时稳定
GPU OOM	Use 4-bit quantization 使用 4-bit 量化
CUDA error	FP16 → load_in_4bit=True

---

📦 Installation / 安装

Option 1: Direct Copy / 方式 1：直接复制

mkdir -p ~/.hermes/skills/data-science/agentic-kaggle/
cp SKILL.md ~/.hermes/skills/data-science/agentic-kaggle/

Option 2: Claude Code CLI / 方式 2：Claude Code CLI

npx skills add FrankS-IntelLab/agentic-kaggle-skill

Option 3: Manual Download / 方式 3：手动下载

Download SKILL.md and place in your skills directory. 下载 SKILL.md 并放入你的 skills 目录。

---

📁 Repository Structure / 仓库结构

agentic-kaggle-skill/
├── SKILL.md                              # Core skill definition / 核心 skill 定义
├── README.md                             # This file / 本文件
├── LICENSE                               # MIT
├── references/research/
│   ├── 01-competition-patterns.md        # Score stabilization, types
│   ├── 02-troubleshooting-guide.md       # Error diagnosis
│   └── 03-automation-patterns.md         # Cronjobs, delegation
└── examples/
    ├── rl-game-case-study.md             # RL game competition
    └── audio-classification-case-study.md # Audio classification

---

🧠 Mental Models / 心智模型

Model 模型	Description 描述
Score Stabilization 分数稳定	Early scores lie — wait 4h for truth 早期分数会骗人
Spec-Driven Development 规范驱动开发	Document before coding, delegate with clarity 先文档后编码
Fail Fast, Learn Faster 快速失败，快速学习	Systematic debugging beats random iteration 系统化调试
Agent as Teammate 智能体即队友	Not just a tool — an autonomous collaborator 不仅是工具

---

🔗 Related Skills / 相关 Skills

Skill	Purpose 用途
agentic-competition-workflow	Git-first project management, validation pipelines Git 项目管理，验证流程
kaggle-auto-submit	End-to-end automation with cronjob 端到端自动化
autonomous-iteration	ANALYSIS → BUILD → EXPERIMENT → REVIEW 分析→构建→实验→审查
opencode	Delegate coding to OpenCode CLI 委托编码
claude-code	Delegate coding to Claude Code CLI 委托编码

---

⭐ Why Star This Repo? / 为什么 Star？

• ✅ Battle-tested patterns from real competitions 实战验证的模式
• ✅ Bilingual documentation (English + 中文) 双语文档
• ✅ Practical troubleshooting for common Kaggle issues 实用故障排除
• ✅ Spec-driven workflow templates included 规范驱动工作流模板
• ✅ Case studies with actual LB scores 带实际 LB 分数的案例研究

---

🤝 Contributing / 贡献

Found a new pattern? Solved a tricky error? 发现了新模式？解决了棘手错误？

1. Fork the repo / Fork 仓库
2. Add your insight to references/research/ or examples/ 添加你的洞察
3. Submit a PR / 提交 PR

---

📄 License / 许可证

MIT — Use freely, modify freely, learn freely. MIT —— 自由使用，自由修改，自由学习。

---

Made with 🤖 by Frank S (IntelLab)

If this skill helped you win a competition, ⭐ star the repo! 如果这个 skill 帮助你赢得了竞赛，请 ⭐ star 这个仓库！