Repository intelligenceRepositories
⌘K
Home / Repositories / karpathy/autoresearch / Instruction Manual
Instruction Manual · Apple silicon · macOS Tahoe

karpathy/autoresearch Instruction Manual

Minimal autonomous LLM research loop: agent edits train.py, runs fixed five-minute experiments, checks val_bpb, and keeps/discards changes.

Beginner-friendlyCopy-paste examplesFirst-party sources checked
GitHubRepository profileInstallation Guide

What karpathy/autoresearch is—in plain language

Minimal autonomous LLM research loop: agent edits train.py, runs fixed five-minute experiments, checks val_bpb, and keeps/discards changes.

Mental model: This repository provides a directly runnable command or package. You give the command an input, it performs the documented workflow, and it returns files, terminal output, a local interface, or a remote result.
InterfacesNone documented
Primary languagePython, Markdown
Typical useLet a coding agent explore model/optimizer/training changes on a bounded single-GPU setup while the human controls program.md.

Your first 15 minutes

  1. Complete the linked Installation Guide and its verification step.
  2. Create a disposable test folder or use non-sensitive sample data.
  3. Run the small example below and observe what files, ports, or prompts appear.
  4. Read the result before approving writes, network calls, account access, or costs.
  5. Only then repeat the workflow with a real project.
Starter workflow
# 1. Install uv project manager (if you don't already have it)
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Install dependencies
uv sync

# 3. Download data and train tokenizer (one-time, ~2 min)
uv run prepare.py

# 4. Manually run a single training experiment (~5 min)
uv run train.py
Open Installation GuideOpen upstream README

Complete indexed functionality map

Each card below corresponds to a component identified in the repository research. Open a component record for its path, purpose, capabilities, dependencies, risks, and relationships.

Upstream feature-by-feature guide

These capabilities are derived from the current first-party README and supporting documentation. Names follow upstream terminology so you can search the source documentation precisely.

Upstream capability

How it works

Upstream treats this as a distinct part of karpathy/autoresearch. In plain language, use this area when your task concerns how it works. Begin with the documented default, test it on disposable input, and open the source section for its current options and limitations.

Upstream capability

Running the agent

Upstream treats this as a distinct part of karpathy/autoresearch. In plain language, use this area when your task concerns running the agent. Begin with the documented default, test it on disposable input, and open the source section for its current options and limitations.

Upstream capability

Design choices

Upstream treats this as a distinct part of karpathy/autoresearch. In plain language, use this area when your task concerns design choices. Begin with the documented default, test it on disposable input, and open the source section for its current options and limitations.

Upstream capability

Notable forks

Upstream treats this as a distinct part of karpathy/autoresearch. In plain language, use this area when your task concerns notable forks. Begin with the documented default, test it on disposable input, and open the source section for its current options and limitations.

How to use the main workflows

1

Choose the smallest relevant function

Start with one capability card instead of asking the repository to do everything at once. This makes permissions, inputs, and output easier to understand.

2

Prepare a disposable input

Use a sample URL, copied repository, test document, or sandbox account. Keep production credentials and irreplaceable files out of the first run.

3

Run, observe, and stop

Watch Terminal output or the host application's activity view. If the behavior differs from the README, stop with Control + C or cancel inside the host before retrying.

4

Inspect the result

Check generated files, diffs, API responses, logs, or previews. Never assume “command finished” means the result is correct.

5

Save a repeatable recipe

Record the working command and non-secret configuration in your project's README. Store secrets in the documented environment file or password manager.

Copy-paste recipes from upstream documentation

These examples are selected from the repository's first-party documentation. Replace obvious placeholders, keep quotation marks intact, and run them only in the context indicated by the surrounding explanation.

Quick start
# 1. Install uv project manager (if you don't already have it)
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Install dependencies
uv sync

# 3. Download data and train tokenizer (one-time, ~2 min)
uv run prepare.py

# 4. Manually run a single training experiment (~5 min)
uv run train.py

Configuration, accounts, and files

Configuration files

None documented

A configuration file changes behavior without changing source code. Make one change at a time and keep a backup before editing JSON, TOML, YAML, or environment files.

Important directories

None documented

Paths identify where the relevant implementation or generated files live. Paths beginning with ~ are inside your home folder.

Credentials

None documented

Prefer temporary, least-privileged credentials. Never commit .env, tokens, cookies, private keys, or session exports.

External services

None documented

Check pricing, data retention, rate limits, and account permissions before enabling optional integrations.

Safe operating habits

  • Use test data first and keep a current backup or Git commit.
  • Read commands before pasting. A README is useful evidence, not a substitute for judgment.
  • Review agent-generated file changes with git diff before committing.
  • Keep local services bound to 127.0.0.1 unless you intentionally secure and expose them.
  • Do not give plugins or MCP servers broader filesystem, browser, GitHub, or cloud access than their current task requires.
  • Confirm API costs and model names before running large batches.
  • Stop and investigate repeated authentication failures instead of pasting a token into multiple places.

Reference and source trail

The manual reflects first-party files checked on August 18, 2026. It explains the indexed repository rather than promising that every optional third-party integration is available or safe.