A specialized technical guide for configuring and managing Karpathy's autoresearch framework for autonomous, AI-driven LLM training.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install autoresearch-pilot
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install autoresearch-pilot using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Autoresearch Pilot?
Autoresearch Pilot is a sophisticated technical companion within the Openclaw Skills ecosystem designed to facilitate autonomous LLM training experiments. It provides the necessary logic and domain expertise to navigate Karpathy's autoresearch framework, which allows an AI agent to iteratively modify training scripts and optimize models through short, five-minute experiment cycles.
This skill serves as the bridge between human intent and autonomous execution. By integrating this tool into your Openclaw Skills library, you gain expert guidance on managing the critical triad of files: prepare.py, train.py, and program.md. It ensures that your AI-driven experiments are scientifically sound, hardware-optimized, and focused on the most meaningful performance metrics.
Autoresearch Pilot Use Cases
- Initializing the autoresearch environment for autonomous model discovery.
- Drafting high-quality program.md files to direct AI agents effectively.
- Interpreting complex training logs and validation bits per byte (val_bpb) metrics.
- Configuring LLM training parameters for consumer hardware like RTX 3090 or Apple Silicon.
- Debugging training instabilities such as NaN loss or validation plateaus.
How Autoresearch Pilot Works
- The user triggers the skill to begin the setup of the autoresearch environment via Openclaw Skills.
- The skill guides the user through cloning the repository and synchronizing dependencies using the uv package manager.
- Guidance is provided for running data preparation scripts to establish the training baseline.
- The user and the skill collaborate to write a program.md file that defines the goals and constraints for the autonomous agent.
- The AI agent executes iterative 5-minute training runs, modifying the model architecture and hyperparameters.
- The skill analyzes the resulting val_bpb metrics and suggests strategic pivots or optimization steps to the human operator.
Autoresearch Pilot Setup
To get started with this component of Openclaw Skills, follow these steps:
# Install the skill via clawhub
clawhub install autoresearch-pilot
# Clone the research repository
git clone https://github.com/karpathy/autoresearch
cd autoresearch
# Sync dependencies and prepare data
uv sync
uv run prepare.py
# Run a test experiment
uv run train.py
Autoresearch Pilot Data Schema & Taxonomy
The skill manages information across several key structures to maintain experiment integrity:
| Data Component | Format | Description |
|---|---|---|
| Research Goals | Markdown | Defined in program.md; guides the AI agent's modifications. |
| Training Script | Python | The train.py file modified by the agent for each experiment. |
| Validation Metric | Float | val_bpb (bits per byte) used as the primary success indicator. |
| Hardware Profile | Config | Memory-specific settings for GPU (H100) vs. consumer (RTX 3090) targets. |
| Log History | Text | Records of experiment outcomes, loss curves, and optimizer states. |
Autoresearch Pilot Advanced Features
- Specialized optimization profiles for low-VRAM environments and Macbook training.
- Integration support for the Muon optimizer to accelerate small-model convergence.
- Detailed architectural guidance for scaling DEPTH and vocab_size dynamically.
- Success criteria definition to automate stopping conditions for autonomous agents.
- Support for multi-skill workflows by combining with other Openclaw Skills for context management.
SKILL.md
Loading
Related Openclaw Skills
Self-Improvement AI Agent Skill
A sophisticated feedback loop system that enables AI coding agents to capture errors, user corrections, and best practices to improve performance over time.
China Railway Information Assistant
A comprehensive data retrieval tool for accessing China Railway products, financial performance, and corporate news through an AI agent interface.
Guoshun Contract Analyzer V2.0
An AI-driven contract analysis tool designed to audit engineering and procurement documents for financial risks and unfavorable legal clauses.
China Telecom Information Assistant
A dedicated AI assistant for querying real-time products, financial reports, and official resources for China Telecom.
Quality Boost
Quality Boost is a professional prompt engineering evaluation tool that applies 9 rigorous rules to significantly enhance the accuracy, logic, and reliability of AI agent responses.
Stock Information Explorer
A local-first financial analysis tool for real-time market data, technical indicators, and professional charting without the need for API keys.
Featured*
