Google Colab

Setup

On first use, read setup.md and align activation behavior and risk boundaries before proposing notebook changes.

When to Use

User needs Google Colab notebook work that must be reproducible, not one-off trial and error. Agent handles runtime setup, package and dependency hygiene, data import and export flows, debugging failures, and experiment tracking.

Architecture

Memory lives in ~/google-colab/. See memory-template.md for setup and status values.

~/google-colab/
|-- memory.md                   # Activation preferences, constraints, and current goals
|-- notebooks.md                # Notebook registry with owners and objective per notebook
|-- runtimes.md                 # Runtime choices, dependency pins, and restart history
|-- datasets.md                 # Data source map, mount paths, and validation notes
|-- incidents.md                # Error timelines, root causes, and fixes
`-- experiments.md              # Hypotheses, metrics, and reproducibility evidence

Quick Reference

Use the smallest relevant file for the active task.

Requirements

• For diagnostics and lightweight API checks: curl, jq
• For notebook execution: Google account with Colab access
• For dataset mounting: explicit permission for Drive, GCS, or external endpoints

Never ask users to paste API keys, OAuth refresh tokens, or private dataset credentials into chat.

Data Storage

Local operational notes stay in ~/google-colab/:

• notebook inventory with objective, owner, and current status
• runtime and dependency decisions with pinned versions
• dataset and schema validation history
• experiment outcomes and unresolved risks

Core Rules

1. Start with Objective, Constraints, and Exit Criteria

Before writing notebook steps, identify:

• objective: prototype, benchmark, fine-tuning, teaching, or production prep
• constraints: runtime tier, budget, execution time, and data availability
• exit criteria: metric threshold, artifact output, or decision checkpoint

Without explicit exit criteria, notebook sessions drift and become hard to evaluate.

2. Design Notebook Cells as Contracts

Each cell should have a contract:

• inputs required and where they come from
• deterministic output shape and validation check
• failure mode and fallback behavior

Treat hidden state between cells as technical debt and document every state dependency.

3. Pin Runtime and Dependency State

Any runnable plan must define:

• Python version and runtime class (CPU, T4, L4, A100 when relevant)
• pinned package versions for non-standard libraries
• rehydration steps after runtime disconnect or restart

Never assume a fresh runtime matches previous package state.

4. Validate Data Paths and Schema Before Training or Evaluation

Before expensive operations:

• verify mount success and path existence
• sample and validate schema and null patterns
• block execution when split boundaries or label columns are ambiguous

Fast schema checks prevent long failed runs and invalid metrics.

5. Make Cost and Time Guardrails Explicit

For any medium or high-cost run:

• estimate runtime duration and checkpoint intervals
• define early-stop conditions and budget cutoff
• recommend smaller dry-run dataset before full execution

No full-scale run should start without budget and cutoff rules.

6. Triage Failures by Layer, Not by Guessing

本技能来自小葱技能站7w4.net。

Debug in layers:

1. runtime health and package import layer
2. data loading and preprocessing layer
3. model logic and training loop layer
4. evaluation and artifact export layer

Layered triage shortens incident resolution and avoids random patching.

7. Log Reproducibility Evidence in Every Significant Run

For each meaningful run, capture:

• notebook id or link, runtime class, seed, and dependency snapshot
• dataset version or timestamp and split method
• primary metric result and whether exit criteria passed

If reproducibility evidence is missing, treat conclusions as provisional.

Common Traps

• Installing packages ad hoc across cells without pins -> results differ after runtime reconnect
• Using absolute local paths copied from old sessions -> file not found during replay
• Training before schema and null validation -> wasted GPU time and misleading metrics
• Mixing exploratory and production cells in one notebook -> brittle execution order
• Treating cached outputs as fresh ground truth -> stale evaluation and wrong decisions
• Ignoring random seeds and data splits -> impossible to compare experiment outcomes
• Exporting artifacts without metadata -> model files cannot be audited later

External Endpoints

No other data should be sent externally unless the user explicitly configures additional systems.

Security & Privacy

Data that leaves your machine:

• notebook payloads and runtime metadata required by Colab services
• selected file and object metadata required for Drive or GCS operations
• package lookup requests for dependency installation

Data that stays local:

• workflow memory and decision logs under ~/google-colab/
• incident notes, experiment summaries, and validation evidence

This skill does NOT:

• request or store raw secrets in conversation text
• execute high-cost runs without explicit guardrails
• bypass user-defined data boundaries or compliance rules

Trust

This skill relies on Google Colab, Google APIs, and package repositories used during notebook setup. Only install and run it if you trust those systems with your code and data.

Related Skills

Install with clawhub install <slug> if user confirms:

• gcp - Plan cloud workloads, storage, and service boundaries for Google environments
• api - Design resilient API contracts for data and model integrations
• pandas - Build robust tabular data transformations and validation pipelines
• numpy - Improve numerical computation patterns and vectorized operations
• automate - Convert repeatable notebook steps into reliable automation workflows

Feedback

• If useful: clawhub star google-colab
• Stay updated: clawhub sync

这是一个设计成熟的 Google Colab 工作流指南，核心规则和调试方法很有实用价值，能帮助建立规范的机器学习实验习惯。安全性方面考虑周全，不会要求用户暴露敏感信息。不足之处是部分内容较浅，缺少实际代码示例，对新手来说可能需要配合其他资源才能真正上手。整体质量良好，但想直接照着操作可能还需要更多具体指引。