modelscope/easydistill

a toolkit on knowledge distillation for large language models

What it solves

EasyDistill 2 is a config-driven toolkit designed to transform black-box teacher models into high-quality training data. It simplifies the process of knowledge distillation by automating the generation and refinement of datasets used for supervised fine-tuning (SFT) and direct preference optimization (DPO).

How it works

The toolkit uses a pipeline-based approach where atomic operators—such as generation, evaluation, filtering, rewriting, balancing, and preference scoring—are composed into end-to-end workflows. These pipelines can be configured via YAML files to produce various types of training data, including text instructions, chain-of-thought (CoT) reasoning, multi-modal conversations, and agent trajectories. It is backend-agnostic, supporting any OpenAI-compatible endpoint (e.g., vLLM, Azure OpenAI) and specialized backends for image and video generation.

Who it’s for

It is intended for developers and researchers who want to create high-quality synthetic datasets to train smaller, specialized models using the knowledge of larger, proprietary or open-source teacher models.

Highlights

  • Multi-modal Support: Distills data for text, vision-language, text-to-image (T2I), and text-to-video (T2V) tasks.
  • Composable Pipelines: Offers both ready-made end-to-end pipelines and individual operators for custom data strategies.
  • Agentic Distillation: Capable of synthesizing virtual tool-use tasks and multi-turn agent trajectories using LangGraph ReAct loops.
  • Training-Ready Outputs: Exports data in formats compatible with LLaMA-Factory and ms-swift.
  • Diverse Data Types: Generates SFT datasets, DPO preference pairs, and complex reasoning traces (CoT).

Related

  • Project
    facebookresearch/stopes

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

  • Project
    yoshitomo-matsubara/torchdistill

    torchdistill is a PyTorch‑based library that lets you run knowledge‑distillation experiments (teacher‑student training) by writing only a YAML configuration. It provides many modern KD losses, a forward‑hook manager for extracting intermediate features, and example scripts for vision and NLP tasks. Install via `pip install torchdistill` and cite the associated papers when used in research.

  • Project
    microsoft/Resource2Skill

    Resource2Skill converts human-created tutorials and multimodal resources into executable skills that AI agents can use to operate professional software like Blender, Excel, and PowerPoint.

  • Project
    titanwings/distilly

    A person-modeling tool that distills a person's experience, voice, and judgment from messages and documents into reusable Person Profiles for AI agents.

  • Project
    facebookresearch/fairseq2

    A modular sequence modeling toolkit for training custom content generation models, supporting large-scale training (70B+ parameters) and high-throughput data pipelines.