The NVIDIA Kaggle Plugin (nvidia-kaggle-skill) empowers AI agents with end-to-end capabilities for Kaggle competitions. It automates complex workflows, enabling agents to gather competition context, analyze top solutions and public kernels, reproduce existing work locally, manage datasets, and submit results. This tool significantly streamlines the process for researchers and power users engaging in data science competitions.

How It Works

This plugin integrates with agent runtimes (like Codex, Claude Code) via a single skill. It interprets natural language requests or specific Kaggle identifiers (URLs, slugs) to execute targeted scripts. The core approach involves abstracting Kaggle API interactions to perform tasks such as fetching competition metadata, summarizing winning strategies from leaderboards, indexing and querying discussions and kernels, downloading notebooks with their dependencies, and facilitating dataset uploads. This unified skill simplifies complex, multi-step Kaggle interactions.

Quick Start & Requirements

• Installation: Add https://github.com/NVIDIA/nvidia-kaggle.git to your agent runtime's plugin marketplace and install nvidia-kaggle. Local installation follows a similar process using . as the marketplace source. For other harnesses, copy the skill directory directly.
• Prerequisites: Python 3.10+, an agent runtime with plugin/skills support, and a Kaggle account with the KAGGLE_API_TOKEN environment variable configured for API-dependent workflows.
• Links:
  • Marketplace/Source: https://github.com/NVIDIA/nvidia-kaggle.git
  • Evaluation Summary: BENCHMARK.md
  • Contributor Guidance: CONTRIBUTING.md

Highlighted Details

• Comprehensive workflow support: competition context, solution writeups, discussions, kernels, kernel reproduction, submission, and dataset management.
• Automated analysis: Summarizes top solutions, researches public kernels for modeling approaches and pitfalls.
• Environment setup: Downloads Kaggle notebooks and their associated data/dependencies into a runnable local workspace.
• Demo available for generating a multi-faceted competition strategy brief.

Maintenance & Community

Contributor guidance is available in CONTRIBUTING.md, and evaluation metrics are detailed in BENCHMARK.md. No specific community channels (e.g., Discord, Slack) or sponsorship information were detailed in the provided README.

Licensing & Compatibility

The project appears to be licensed under Apache 2.0, as indicated in an example output. Compatibility is primarily focused on agent runtimes that support the Agent Skills or plugin architecture.

Limitations & Caveats

Reproducibility can be impacted if referenced datasets within kernels are private or deleted, as noted in an example. Core functionalities heavily rely on a correctly configured KAGGLE_API_TOKEN. The effectiveness is contingent on the agent runtime's integration capabilities.