<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> EasyDistill is a pioneering toolkit for knowledge distillation (KD) of large language models (LLMs), enabling smaller models to emulate larger ones efficiently. It targets NLP researchers and practitioners, offering a versatile, user-friendly platform to streamline KD, support diverse methodologies, and facilitate practical industrial solutions.

How It Works

The toolkit supports black-box and white-box KD, featuring data synthesis, SFT, ranking optimization, and RL. It accommodates System 1 (fast, intuitive) and System 2 (slow, analytical) cognitive models via a modular architecture and simple CLI, facilitating experimentation and industrial integration with platforms like Alibaba Cloud PAI.

Quick Start & Requirements

Clone the repo (git clone https://github.com/modelscope/easydistill), navigate (cd EasyDistill), and install with python setup.py install. Run jobs via easydistill --config <config-file-path>. Specific hardware requirements like GPU memory are configurable within job settings.

Highlighted Details

• DistilQwen Series: Distilled LLMs from Qwen models, offering reduced size with high performance for resource-constrained environments.
• Adaptive Thinking Models: DistilQwen-ThoughtX/Y models feature improved reasoning, using the OmniThought dataset with RV/CD scores for optimal CoT generation.
• System 1 & 2 Models: DistilQwen2/2.5 enhance instruction following. DistilQwen2.5-R1/DS3 focus on reasoning, using DeepSeek-R1 as a teacher and employing CogPO/CoT simplification.
• Public Datasets: Includes released instruction-following (DistilQwen_100K/1M) and Chain-of-Thought (OmniThought, OmniThought-0528) datasets, annotated with RV/CD scores.

Maintenance & Community

The project actively releases new models and features, with recent updates including OmniThoughtV and multi-modal KD. Community discussions are welcomed via a DingTalk group (117440002081). Integration with Alibaba Cloud PAI is supported.

Licensing & Compatibility

The primary license is Apache License 2.0, permissive for commercial use. However, some included code may originate from other repositories under different licenses; consult the NOTICE file for specifics.

Limitations & Caveats

The README does not explicitly detail known limitations or bugs. Users should review the NOTICE file for potential licensing complexities from incorporated third-party code.