Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
a toolkit on knowledge distillation for large language models
| Date | Stars |
|---|---|
| 2026-07-31 | 443 |
| 2026-08-05 | 445 |
| 2026-08-06 | 445 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# EasyDistill: Easy Knowledge Distillation for Large Language Models <div align="center"> [中文](./README_zh.md) | [English](./README.md) </div> Introducing **EasyDistill**, a pioneering toolkit on knowledge distillation (KD) for large language models (LLMs). With the growing complexity and size of LLMs, **EasyDistill** offers a versatile and user-friendly platform to streamline the KD process, supporting both black-box and white-box methodologies. It facilitates efficient model training, enabling smaller models to emulate the performance of larger ones without compromising accuracy. **EasyDistill** boasts an extensive range of features, including data synthesis, supervised fine-tuning, ranking optimization, and reinforcement learning, all tailored for various KD scenarios. Designed to accommodate both System 1 (fast, intuitive) and System 2 (slow, analytical) cognitive models, the toolkit is modular and easy to use, with a simple command-line interface guiding users. Beyond academic exploration, **EasyDistill** anchors practical industrial solutions, offering robust distilled models and open-source datasets, while also showcasing seamless integration with Alibaba Cloud’s AI platform, PAI. Committed to bridging theoretical advancements with practical needs, **EasyDistill** empowers the NLP community, making state-of-the-art KD strategies accessible to researchers and industry practitioners alike. # News - **AgentKD**: We have released AgentKD, which can **directly generate diverse virtual tool-use tasks** from persona seeds to tools, policies and test cases, and produce teacher-model solution trajectories for knowledge distillation. Virtual tools are **pre-defined** with their schemas and possible return results; the **LLM simulates** tool execution (no real API calls). The data can also be used for **RL training**, with evaluation rubrics provided. Refer to [Here](./easydistill/agentkd). - Dec 19th: We have released OmniThoughtV, multi-modal CoTs distilled and filtered from Qwen-VL-max. - July 28th: We have released the functionalities of knowledge distillation from MLLM (aka MMKD). Refer to [Here](./easydistill/mmkd). Evaluations on the qualities of instruction-following and CoT datasets have been updated. Refer to [Here](./easydistill/eval). - June 25th: We have released a new series of DistilQWen models named DistilQwen-ThoughtY, togeter with OmniThought-0528 (CoTs distilled from DeepSeek-R1-0528). # Technical Articles We have a series of technical articles on the functionalities of EasyDistill. - [面向 Interleaved Thinking 的大模型 Agent 蒸馏实践](https://zhuanlan.zhihu.com/p/1992286983679141168) - [OmniThoughtV：面向多模态深度思考的高质量数据蒸馏](https://mp.weixin.qq.com/s/DoIpiauTxLwR9TtaZwbK5w) - [基于模型蒸馏的大模型文案生成最佳实践](https://developer.aliyun.com/article/1675249) - [DistillQwen-ThoughtY：通过变长思维链蒸馏，全面提升模型推理能力！](https://developer.aliyun.com/article/1669748) - [DistilQwen-ThoughtX：变长思维链推理模型，能力超越DeepSeek蒸馏模型](https://developer.aliyun.com/article/1665220) - [阿里云人工智能平台 PAI 开源 EasyDistill 框架助力大语言模型轻松瘦身](https://developer.aliyun.com/article/1664823) - [人工智能平台 PAI DistilQwen2.5-DS3-0324发布：知识蒸馏+快思考=更高效解决推理难题](https://developer.aliyun.com/article/1661734) - [DistilQwen2.5-R1发布：知识蒸馏助推小模型深度思考](https://developer.aliyun.com/article/1659288) - [DistilQwen2.5发布：通义千问蒸馏小模型再升级](https://developer.aliyun.com/article/1653842) - [DistilQwen2：通义千问大模型的知识蒸馏实践](https://developer.aliyun.com/article/1633882) - [基于多轮课程学习的大语言模型蒸馏算法TAPIR](https://developer.aliyun.com/article/1635146) ## Overview  - **Toolkit Features**: EasyDistill provides versatile functionalities, including data synthesis, supervised fine-tuning, logits distillation, ranking optimization, and reinforcement learning techniques tailored for KD scenarios. **AgentKD** extends the toolkit with virtual tool-use data synthesis: tools and their possible return results are pre-defined; the LLM simulates tool execution. It generates diverse tool-use metadat
Excerpt of 22,963 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:bbb255e245af8ec2, topic:knowledge-distillation, desc:knowledge distillation
matched fp:bbb255e245af8ec2, topic:large-language-models