thinking-machines-lab/tinker-cookbook

Post-training with Tinker

What it solves

It simplifies the process of customizing and fine-tuning large language models (LLMs) by removing the complexities of managing distributed training infrastructure. It provides a high-level SDK and a library of practical recipes for various post-training techniques.

How it works

The project consists of two main components: tinker, a training SDK that allows users to send API requests to a managed service for distributed training (handling operations like forward/backward passes and optimizer steps), and tinker-cookbook, which provides high-level abstractions and concrete implementation examples (recipes) built on top of that API.

Who it’s for

It is designed for researchers and developers who want to fine-tune LLMs using techniques like SFT, RL, and DPO without managing their own GPU clusters.

Highlights

  • Diverse Training Recipes: Includes ready-to-use examples for Chat SFT, Math and Code RL, Preference Learning (DPO/RLHF), Knowledge Distillation, Tool Use (RAG), and Multi-agent self-play.
  • Multimodal Support: Supports fine-tuning for audio and image inputs, specifically for the Inkling model family.
  • Integrated Evaluation: Features a benchmark framework supporting 12+ standard benchmarks (e.g., GSM8K, MMLU-Pro) to evaluate trained models.
  • Claude Code Integration: Provides specialized skills for the Claude Code agent to help users plan, debug, and implement training experiments.

Related

  • Project
    ZHZisZZ/dllm

    A library that unifies the training, inference, and evaluation of diffusion language models, providing scalable pipelines and recipes for converting AR models into diffusion models.

  • Project
    agentscope-ai/Trinity-RFT

    A general-purpose framework for reinforcement fine-tuning of LLMs that decouples exploration, training, and data buffering to support diverse RL algorithms and agentic workflows.

  • Project
    NovaSky-AI/SkyRL

    A modular full-stack reinforcement learning library for LLMs, providing tools for training long-horizon, multi-turn agents capable of tool use in real-world environments.

  • Project
    google/tunix

    A JAX-based post-training library for LLMs that provides scalable support for supervised fine-tuning and reinforcement learning, specifically optimized for TPU performance.

  • Project
    datawhalechina/diy-llm

    A Chinese‑language, open‑source curriculum that teaches you to build, train, and evaluate large language models from scratch, complete with lecture notes, six hands‑on assignments, and runnable PyTorch code.