Tinker Cookbook

Tinker Cookbook is a practical collection of examples for the Tinker platform, aimed at researchers and engineers building reproducible training pipelines.

Detailed Introduction

Tinker Cookbook is a repository maintained by Thinking Machines Lab that demonstrates how to use the Tinker platform for model training and post-processing. The repository includes examples ranging from supervised fine-tuning to reward-based reinforcement learning, multi-agent setups, and tool-use training pipelines. Each subfolder contains a README with environment setup, required API key configuration, example datasets, and expected outputs to facilitate reproducible experiments. The examples reference Large Language Models (LLM) where applicable.

Main Features

• Rich examples: covers supervised fine-tuning, RLHF, preference learning, prompt distillation, and multi-task training.
• Reproducibility: each example includes run instructions and minimal dependencies for quick setup using pip install -e ..
• Platform integration: shows how to use Tinker training and sampling clients to run training, save checkpoints, and download artifacts via API.
• Open license: released under Apache-2.0, enabling reuse in research and engineering.

Use Cases

• Research reproduction: quickly reproduce fine-tuning and evaluation workflows from papers.
• Prototyping: build small-scale training pipelines to validate method ideas.
• Teaching and workshops: serve as hands-on material to illustrate fine-tuning and evaluation practices.
• Engineering integration: adapt existing training logic into managed remote training jobs using the Tinker API.

Technical Features

• Training-oriented utilities: includes helper functions and scripts for learning rate configuration, LoRA setup, and hyperparameter calculations.
• Modular organization: recipes are organized by task for easy composition and extension.
• Evaluation support: contains evaluation abstractions and integration examples to measure model performance on standard benchmarks.
• Environment notes: provides example data, dependency specifications (pyproject.toml), and developer guides to reduce onboarding time.