Tinker Cookbook - Tinker Documentation
Tinker Cookbook
The Tinker Cookbook is the ecosystem around the Tinker SDK. While the SDK provides the core client-server API for training and sampling, the cookbook gives you everything else:
- Training abstractions —
Env,EnvGroupBuilder,RLDataset,train.Config— so you can focus on your task, not plumbing - Renderers — Convert chat messages to tokens for every supported model family
- Production recipes — 15+ end-to-end training examples (SFT, RL, DPO, distillation, tool use, multi-agent, and more)
- Weight management — Download, merge, export, and publish trained models to HuggingFace
- Evaluation — 20+ standardized benchmarks (GSM8K, MMLU, MATH, IFEval, SWE-Bench, …) with concurrency, trajectory storage, and training-loop integration
- Using Inkling — Render chat, tool calls, images, and audio with
tml-renderers
When to use the Cookbook vs. the SDK
| Tinker SDK | Tinker Cookbook | |
|---|---|---|
| What | Core API client | Training ecosystem |
| Use when | You want full control over every API call | You want batteries-included training |
| Examples | training_client.forward_backward(data, "cross_entropy") |
asyncio.run(train.main(config)) |
| Install | uv pip install tinker |
uv pip install tinker-cookbook |
Most users will use both — the SDK for understanding what's happening, and the cookbook for building production training pipelines.
Getting Started
- Install the cookbook
- Follow the Tutorials — they cover both SDK and cookbook
- Browse the Recipes for production examples
- Check the API Reference when you need details