Guide for developing and contributing to tinker-cookbook.

The cookbook uses the builder pattern for datasets: a `*DatasetBuilder` (config) builds a `*Dataset` (runtime).

Set up and run knowledge distillation (on-policy, off-policy, or multi-teacher) from a teacher model to a student model using the Tinker API. Use when the user wants to distill knowledge, compress models, or train a student from a teacher.

Set up and run Direct Preference Optimization (DPO) training on preference datasets using the Tinker API. Use when the user wants to train with preference data, chosen/rejected pairs, or DPO.

Set up and run reinforcement learning with verifiable rewards (RLVR/GRPO) for math, code, or custom environments using the Tinker API. Use when the user wants to do RL training, GRPO, reward-based optimization, or train with verifiable rewards.

Guide for training outputs, metrics logging, logtree reports, tracing/profiling, and debugging training runs. Use when the user asks about training logs, metrics, debugging, tracing, profiling, timing, Gantt charts, or understanding training output files.

Create, update, or organize Claude Code skills in this repo. Use when adding a new skill, reviewing existing skills for consistency, or maintaining the skill taxonomy.

Help the user choose the right model for their task.

Set up and run multi-turn RL training for interactive environments (terminal tasks, tool use, search/RAG, games) using the Tinker API. Use when the user wants multi-turn RL, agentic training, tool-use RL, or interactive environment training.

Renderers convert chat-style messages into token sequences for training and generation.

Set up and run the full RLHF pipeline (SFT, reward model training, RL from reward model) using the Tinker API. Use when the user wants to do RLHF, train a reward model, or run the full preference-based RL pipeline.

Set up and run supervised fine-tuning (SFT) on instruction or chat datasets using the Tinker API. Use when the user wants to do instruction tuning, chat fine-tuning, or supervised learning.

The repo has two layers of testing and two CI workflows.

The `tinker` CLI is installed with the Tinker Python SDK. It provides commands for managing training runs and checkpoints from the terminal.

Quick reference for the core types used throughout the Tinker SDK and cookbook.

The `tinker_cookbook.weights` subpackage provides a standard pipeline for trained weight management: **download → build → publish**.