AI Makerspace - Tinker API Workshop

What is Tinker?

Tinker is an API from Thinking Machines Lab that provides direct access to LLM training primitives. Unlike inference-only APIs, Tinker lets you fine-tune and train language models programmatically.

Tinker enables you to:

• Access low-level training operations (forward_backward(), optim_step(), sample(), save_state())
• Fine-tune models using LoRA without managing infrastructure
• Experiment with custom training loops and reward functions
• Build and iterate on ML research ideas quickly

About This Workshop

This hands-on workshop introduces the Tinker API through a practical demo. You'll learn how to use Tinker's training primitives to improve a language model's capabilities in real-time.

What You'll Learn

Through practical examples in the guess_number_demo.ipynb, you'll explore:

1. Tinker's Core API - Understanding the training primitives
2. Sampling - Generating outputs from your model
3. Training Steps - Running forward/backward passes and optimizer steps
4. Custom Training Loops - Building your own training logic

Why This Matters

Tinker democratizes access to LLM training by providing:

• Accessibility - Train models without managing GPU clusters
• Flexibility - Full control over the training process
• Experimentation - Rapid iteration on training ideas
• Education - Learn how LLM training actually works

Whether you're a researcher, hobbyist, or developer, Tinker opens up LLM training in a way that wasn't previously accessible.

Setup Instructions

1. Install Dependencies

uv sync

2. Get Your Tinker API Key

1. Sign up at auth.thinkingmachines.ai/sign-up
2. Verify your email - Check your inbox for a verification link
3. Log in to the Tinker Console
4. Create an API key - Navigate to the API Keys section and generate a new key
5. Save your key - Copy the key immediately (you won't be able to see it again)

3. Configure Your Environment

Create a .env file in the project root:

echo "TINKER_API_KEY=your_key_here" > .env

Or manually create .env with:

TINKER_API_KEY=your_api_key_here

4. Run the Demo

uv run jupyter lab guess_number_demo.ipynb

Workshop Structure

The notebook is organized into focused sections:

• Setup & Configuration - API keys and environment preparation
• Understanding Tinker - Overview of the API and its primitives
• Sampling - Generating model outputs
• Training - Running training steps with custom rewards
• Evaluation - Measuring model improvement

Key Takeaways

By completing this workshop, you'll understand how to:

• Connect to and authenticate with Tinker's API
• Use sample() to generate model outputs
• Run forward_backward() and optim_step() for training
• Build custom training loops for your own use cases
• Apply Tinker to your own ML projects

Resources

• Tinker Documentation
• Tinker Cookbook - More recipes and examples
• Thinking Machines Lab

Files

• guess_number_demo.ipynb - Main demo notebook
• main.py - Supporting Python code
• .env - Your API key (not tracked in git)

Start exploring Tinker by opening the notebook and running through the examples!