thinking-machines-lab/tinker-cookbook

Post-training with Tinker

What it solves

Tinker Cookbook provides a set of high-level abstractions and realistic examples to simplify the process of customizing and fine-tuning large language models (LLMs). It removes the complexities of distributed training by leveraging the Tinker API, allowing developers to focus on training recipes rather than infrastructure.

How it works

The library builds on the tinker SDK, which handles the distributed training backend. Tinker Cookbook provides a library of "recipes"—pre-configured training environments—and utilities for rendering token sequences into chat messages, managing hyperparameters for LoRA training, and evaluating models.

Who it’s for

It is designed for researchers and developers who want to perform advanced post-training techniques like Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), and Distillation without managing their own training clusters.

Highlights

  • Diverse Training Recipes: Includes implementations for Chat SFT, Math and Code RL (with sandboxed execution), Preference Learning (DPO/RLHF), Knowledge Distillation, and Multi-agent RL.
  • Multimodal Support: Supports SFT and RL for audio and image inputs, specifically for the Inkling model family.
  • Integrated Evaluation: An experimental benchmark framework supporting 12+ benchmarks (e.g., GSM8K, MMLU-Pro, GPQA) to verify model performance.
  • Claude Code Integration: Ships with specialized skills to help AI assistants automate the planning and debugging of 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 modes and agentic workflows.

  • Project
    huggingface/alignment-handbook

    A collection of robust training recipes and scripts to continue pretraining and align language models with human and AI preferences using techniques like DPO and ORPO.

  • 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.

  • Project
    facebookresearch/stopes

    A library for preparing machine translation data, providing tools for cleaning monolingual data, mining bitext, and running knowledge distillation pipelines.