Shaping capabilities with token-level pretraining data filtering

Our code for ICLR'25 paper "DataMan: Data Manager for Pre-training Large Language Models".

Code implementation of synthetic continued pretraining

A compact implementation of SGLang, designed to demystify the complexities of modern LLM serving systems.

repo for paper https://arxiv.org/abs/2504.13837

一个基于nano banana pro🍌的原生AI PPT生成应用，迈向真正的＂Vibe PPT＂; 支持上传任意模板图片；上传任意素材&智能解析；一句话/大纲/页面描述自动生成PPT；口头修改指定区域、一键导出可编辑ppt - An AI-native PPT generator based on nano banana pro🍌

Code for VideoCompressa: Data-Efficient Video Understanding via Joint Temporal Compression and Spatial Reconstruction

Crowdfunding open source projects: use OpenReview's high-quality review data to fine-tune a professional review and response LLM. 众筹开源项目：利用OpenReview的优质审稿数据，微调出一个专业的审稿和审稿回复GPT

Agent framework and applications built upon Qwen>=3.0, featuring Function Calling, MCP, Code Interpreter, RAG, Chrome extension, etc.

Tools for merging pretrained large language models.

The official implementation for [NeurIPS2025 Oral] Gated Attention for Large Language Models: Non-linearity, Sparsity, and Attention-Sink-Free

Enjoy the magic of Diffusion models!

The official implementation of dLLM-Var

Physics of Language Models: Part 4.2, Canon Layers at Scale where Synthetic Pretraining Resonates in Reality

Search Self-Play: Pushing the Frontier of Agent Capability without Supervision

Scaling Preference Data Curation via Human-AI Synergy

Code repository for Group-MATES Group-Level Data Selection for Efficient Pretraining

Socratic-Zero is a fully autonomous framework that generates high-quality training data for mathematical reasoning

This is the official implementation for Human-Agent Collaborative Paper-to-Page Crafting for Under $0.1.

ERGO (Efficient Reasoning & Guided Observation) is a large vision–language model trained with reinforcement learning on efficiency objectives.