Tinker Cookbook

Tinker Cookbook

Tinker Cookbook (this repository) is a comprehensive library designed to empower researchers and developers to fine-tune large language models (LLMs) using the Tinker API from Thinking Machines Lab. It builds upon two core libraries: tinker, a training SDK that abstracts away the complexities of distributed training via simple API requests, and tinker-cookbook, which provides high-level abstractions and realistic examples for various fine-tuning scenarios.

Key Features and Primitives

The tinker SDK offers essential primitives for LLM fine-tuning:

• Training Client: Create LoRA training clients (e.g., on base models like Llama-3.2-1B), perform forward-backward passes, optimization steps, save/load states.
• Sampling Client: Generate samples from trained models.
• Checkpoint Management: Download model weights via REST API.

Example usage is shown in minimal loops for supervised learning (sl_loop.py) and reinforcement learning (rl_loop.py).

Cookbook Recipes

The cookbook includes a diverse set of sophisticated examples in the recipes/ folder:

1. Chat Supervised Learning: Fine-tune on conversational datasets like Tulu3 (chat_sl/).
2. Math Reasoning: Reward correct math answers to boost reasoning (math_rl/).
3. Preference Learning: Full RLHF pipeline—SFT, reward model training, RL (preference/).
4. Tool Use: Enhance tool retrieval for accurate Q&A (tool_use/).
5. Prompt Distillation: Internalize complex instructions (prompt_distillation/).
6. Multi-Agent: Optimize LLMs for self-play or vs. other LLMs (multiplayer_rl/).

Each recipe has a detailed README.md with implementation, run commands, and performance expectations. Start with basics in sl_basic.py and rl_basic.py.

Utilities

• renderers.py: Convert tokens to/from structured chat messages.
• hyperparam_utils.py: Compute LoRA hyperparameters.
• evaluation/: Abstractions for model eval; integrates with InspectAI for benchmarks (inspect_evaluation.py).

Getting Started

1. Join Tinker waitlist: thinkingmachines.ai/tinker.
2. Get API key from console, set TINKER_API_KEY.
3. pip install tinker and pip install -e . (in venv).
4. See docs for basics.

Community and Citation

Open for PRs post-private beta. Feedback: tinker@thinkingmachines.ai. Cite as:

@misc{tml2025tinker,
  author = {Thinking Machines Lab},
  title = {Tinker},
  year = {2025},
  url = {https://thinkingmachines.ai/tinker/},
}

Launched in October 2025 as Thinking Machines Lab's first product, Tinker Cookbook has garnered over 2.3k stars, enabling efficient, API-driven LLM customization without infrastructure hassles.