Projects

### Overview: Infinity is a cutting-edge visual synthesis framework that leverages bitwise auto-regressive modeling to generate high-resolution and photorealistic images. This project is a significant advancement in the field of visual autoregressive models, achieving state-of-the-art performance on various benchmarks. ### Key Features: - **Bitwise AutoRegressive Modeling**: Uses a bitwise token prediction framework with an infinite-vocabulary tokenizer and classifier for powerful scaling capabilities. - **Infinite Vocabulary Tokenizer**: Enables the model to scale vocabulary size to infinity while maintaining high performance. - **Infinite Transformer**: Utilizes an infinite transformer to further enhance the model's capabilities. - **Generative Performance**: Outperforms top-tier diffusion models like SD3-Medium and SDXL in terms of accuracy and efficiency. - **Benchmark Improvements**: Achieves a win rate of 66% in the GenEval benchmark, surpassing SD3-Medium by 0.11 points and SDXL by 0.11 points. ### Tech Stack: - **Python**: The primary programming language used for implementing the project. ### Why It Matters: Infinity represents a significant leap forward in visual autoregressive modeling, offering unparalleled performance and scalability. Its ability to generate high-resolution images with photorealistic details makes it a valuable tool for various applications, including art, design, and scientific visualization. ### Repository: - **GitHub Repository**: [FoundationVision/Infinity](https://github.com/FoundationVision/Infinity) - **GitHub Stars**: 1593 - **GitHub Description**: [CVPR 202

Overview: Functionary is a language model capable of interpreting and executing functions, providing a seamless function calling experience. Key features: - Function calling with JSON Schema Objects - Ability to execute functions in parallel or serially - Understands function outputs - Triggers functions only as needed Tech stack: Python Why it matters: Functionary simplifies the process of integrating external tools into a chat interface, making it easier for developers to leverage external functionalities without the complexity of writing scripts.

### Overview: SGLang-Omni is a high-performance serving framework designed for audio models (TTS, ASR) and unified multimodal models. Built with Python, it offers fast and efficient processing of audio and multimodal data. ### Key features: - **Unified Multimodal Processing**: Seamlessly integrates audio and text inputs for advanced multimodal applications. - **Fast Inference**: Utilizes optimized engines and pipelines to deliver high-performance inference. - **Streamed Outputs**: Supports real-time streaming of audio and text outputs, ideal for live applications. ### Tech stack: Python ### Why it matters: SGLang-Omni is crucial for developers looking to build robust and scalable systems that can handle complex audio and multimodal processing tasks. Its ability to integrate seamlessly with various models and platforms makes it an essential tool in the development of intelligent applications.

### Overview: AIDE ML is an open-source Python package designed to automate the writing, debugging, and benchmarking of machine learning code. It is a research-friendly implementation of the AIDE algorithm, which is a tree-search agent that optimizes metrics using LLMs. ### Key Features: - **Natural-Language Task Specification**: Easily describe goals and metrics in plain English. - **Iterative Agentic Tree Search**: Each script becomes a node in a solution tree, and LLM-generated patches spawn children. The algorithm outperforms linear agents in complex tasks like MLE-Bench. - **HTML Visualiser**: Inspect the full solution tree and code attached to each node. - **Streamlit UI**: Prototype ML solution. - **Model-Neutral Plumbing**: Supports OpenAI, Anthropic, Gemini, or any local LLM that speaks the OpenAI API. ### Tech Stack: Python ### Why It Matters: AIDE ML is a research-friendly Python package that enables academics and engineers to replicate the results of the paper, test new ideas, and prototype ML pipelines. It outperforms linear agents in complex tasks like MLE-Bench, demonstrating the effectiveness of the AIDE algorithm. ### Repository: Repository: WecoAI/aideml GitHub Stars: 1591 GitHub Description: AIDE: an LLM agent for machine learning engineering - the research Weco grew out of. Referenced in OpenAI MLE-bench.

### Overview: USB is a unified semi-supervised learning benchmark designed for computer vision, natural language processing, and audio classification. It provides a comprehensive set of datasets and models to evaluate the effectiveness of different semi-supervised learning techniques. ### Key Features: - **Unified Benchmark**: USB offers a single, unified environment for evaluating semi-supervised learning across various domains. - **Datasets**: Includes datasets such as CIFAR-10, CIFAR-100, MNIST, and more. - **Models**: Supports a variety of models including ResNet, DenseNet, and transformers. - **Evaluation Metrics**: Utilizes metrics like accuracy, F1 score, and AUC to assess model performance. ### Tech Stack: Python ### Why it Matters: USB provides a standardized benchmark that helps researchers and practitioners evaluate the performance of semi-supervised learning techniques across different domains. It serves as a valuable resource for advancing the field and comparing the strengths and weaknesses of various semi-supervised methods. Repository: microsoft/Semi-supervised-learning GitHub stars: 1589 GitHub description: A Unified Semi-Supervised Learning Codebase (NeurIPS'22) GitHub URL: [https://github.com/microsoft/Semi-supervised-learning](https://github.com/microsoft/Semi-supervised-learning)

### Overview: Agentlas OS is a decentralized agent hub that allows specialists to be easily summoned and managed. It works seamlessly with any model and can be used with Claude Code, Codex, Cursor, and Antigravity, as well as local models. ### Key Features: - **Agent Summoning:** Specialized agents can be called up instantly as needed. - **Model Compatibility:** Works with a variety of models, including Claude Code, Codex, Cursor, and Antigravity. - **Local-first Approach:** Utilizes local models for faster performance and reduced latency. ### Tech Stack: Python ### Why it Matters: Agentlas OS simplifies the process of bringing in specialized talent, making it easier to find the right specialist for any project. It's designed to be a reliable and efficient solution for teams looking to enhance their capabilities.

Overview: This project is an open-source implementation of a poker AI using the ideas of Counterfactual Regret. Key features: - Cluster Hero Information - Train your bot - Play your bot Tech stack: Python Why it matters: This project provides a best-effort implementation of a poker AI, showcasing the application of ideas from Counterfactual Regret.

Overview: This framework provides a comprehensive set of tools for creating AI motion in Godot games, supporting both 2D and 3D environments, and includes various steering behaviors for smooth and efficient movement. Key features: - Steering behaviors for AI motion - Support for 2D and 3D games - Includes flee, follow, look at, and blended behaviors - Works seamlessly with the Godot game engine Tech stack: GDScript Why it matters: This framework simplifies the process of creating complex AI movements in games, making it easier to implement sophisticated AI behaviors without the need for complex pathfinding algorithms.

### Overview: Fun-ASR is a family of end-to-end speech recognition models from Tongyi Lab. It includes models for Chinese, English, Japanese, and various dialects and accents. The models are trained on tens of millions of hours of speech data and support multiple languages and accents. ### Key features: - **Chinese Support**: Supports Chinese dialects (Wu, Cantonese, Min, Hakka, Gan, Xiang, Jin) and regional accents (Henan, Shanxi, Hubei, Sichuan, Chongqing, Yunnan, Guizhou, Guangdong, Guangxi, and more than 20 other regions). - **English and Japanese Support**: Covers multiple regional accents. - **Lyric Recognition and Rap Speech Recognition**: Additional features for enhanced speech recognition. - **Training Data**: Trained on tens of millions of hours of speech data. ### Tech stack: Hugging Face Transformers ### Why it matters: Fun-ASR provides comprehensive speech recognition capabilities, including support for multiple languages and dialects, making it suitable for a wide range of applications and environments.

Overview: FIM One is an all-in-one agent platform designed for global enterprises, enabling seamless integration of their existing systems into a unified AI core. Key features: - Wire every system they run, bridging global SaaS solutions with the full China stack (Feishu, WeCom, DingTalk, DM, Kingbase, etc.). - Supports both standalone and hub modes, providing flexibility in deployment. - Offers dashboard for monitoring, quick access to agents and conversations, and real-time agent chat functionality. Tech stack: Python, Docker, Node.js, pnpm, and various LLMs. Why it matters: FIM One simplifies AI integration, enabling enterprises to leverage AI across their global and China-based systems seamlessly, enhancing productivity and efficiency. Repository: fim-ai/fim-one GitHub stars: 1585 GitHub description: Open-source agent platform for Global × China enterprises — wire every system through one agent core. Self-hosted, any LLM. Tech stack: Python, Docker, Node.js, pnpm, and various LLMs.

### Overview: DiffDock is an implementation of the state-of-the-art method for molecular docking, designed by Gabriele Corso, Hannes Stark, Bowen Jing, Regina Barzilay, and Tommi Jaakkola. This project focuses on improving the performance and generalization capacity of molecular docking through the Diffusion Steps, Twists, and Turns method. ### Key features: - **Diffusion Steps, Twists, and Turns Method**: This feature is the core of the project, enhancing the accuracy and generalization of molecular docking simulations. - **Molecular Docking Implementation**: The repository provides the necessary code and instructions to run the DiffDock method effectively. - **Performance Improvement**: The latest version, DiffDock-L, offers significant improvements in performance and generalization capacity compared to the original DiffDock. ### Tech stack: Python ### Why it matters: This project is crucial for the field of molecular docking, offering advanced techniques that can significantly enhance the accuracy and reliability of docking simulations.

Overview: This repository provides a high-quality, fast, and modular reference implementation of SSD (Single Shot MultiBox Detector) in PyTorch, aiming to serve as a code base for various researches based on SSD. Key features: - PyTorch 1.0 support - Multi-GPU training and inference - Modular design allowing easy replacement of components - CPU support for inference - Batched inference - Evaluating during training - Metrics visualization in TensorBoard - Auto download of pre-trained weights Tech stack: Python, PyTorch 1.0, yacs, OpenCV, Vizer Why it matters: This repository offers a comprehensive and efficient way to implement SSD, making it a valuable resource for researchers and developers looking to build and evaluate SSD models.

### Overview: **similarity** 是一个基于Java的文本相似度计算工具包，旨在简化自然语言处理中的相似度计算。该工具包提供了一系列算法来计算词语、短语和句子的相似度，适用于文本相似度计算、情感分析等任务。 ### Key features: - **词语相似度计算**：支持词林编码法、汉语语义法、知网词语相似度和字面编辑距离法。 - **短语相似度计算**：采用简单短语相似度和词性和词序结合法。 - **句子相似度计算**：提供词性和词序结合法、编辑距离算法、Gregor编辑距离法、优化编辑距离法。 - **段落相似度计算**：支持余弦相似度、编辑距离、欧几里得距离、Jaccard相似性系数、Jaro距离、Jaro–Winkler距离、曼哈顿距离和SimHash + 汉明距离。 - **知网义原**：提供词语义原树。 - **情感分析**：支持正面倾向程度、负面倾向程度和情感倾向性。 - **近似词**：使用word2vec进行近似词处理。 ### Tech stack: Java ### Why it matters: **similarity** 是一个开箱即用的工具包，适用于文本相似度计算、情感分析等任务，简化了自然语言处理中的相似度计算过程，提高了开发效率。

Overview: Bundler is a structure-from-motion system designed for processing unordered image collections, such as those from the Internet, to reconstruct a 3D scene and sparse scene geometry. Key features: - Supports unordered image collections - Incremental reconstruction of the scene - Uses Sparse Bundle Adjustment for optimization - Compatible with Linux, Windows (Cygwin), and Visual Studio Tech stack: C Why it matters: Bundler is a powerful tool for analyzing and reconstructing 3D scenes from digital images, making it valuable for applications in computer vision, archaeology, and more. Repository: snavely/bundler_sfm GitHub stars: 1579 GitHub description: Bundler Structure from Motion Toolkit

Overview: MIMO is an official PyTorch implementation of "MIMO: Controllable Character Video Synthesis with Spatial Decomposed Modeling". This project aims to provide a generalizable model for controllable video synthesis, enabling the synthesis of realistic character videos with controllable attributes such as character, motion, and scene, even with very simple user inputs. Key features: - Realistic character video synthesis - Controllable attributes (character, motion, scene) - Achieves advanced scalability and generality - Supports interactive real-world scenes Tech stack: Python, PyTorch, TensorFlow, CUDA 12.1, GPU (tested on A100, L20) Why it matters: MIMO offers a unified framework for controllable video synthesis, making it a valuable tool for researchers and developers looking to create more interactive and realistic video content.

### Overview: Youwee is a beautiful, cross-platform downloader for YouTube, TikTok, Instagram, and 1800+ sites, offering AI video summaries and post-processing features. It's a free and open-source project that uses TypeScript and supports multiple languages, including English, Vietnamese, and Chinese. ### Key features: - **Video Downloads**: Supports YouTube, TikTok, Facebook, Instagram, Bilibili, Youku, and over 1800+ sites. - **Browser Extension Bridge**: Chromium + Firefox extension with a floating button, media/quality picker, and download options. - **Remote Download**: Control downloads remotely with Telegram commands. - **Plugins & Workflow Automation**: Install plugins, configure custom fields, and extend Youwee with notifications, uploads, and post-download automations. - **Channel Follow**: Follow YouTube, Bilibili, and Youku channels, get notifications, and auto-download. - **Metadata Fetcher**: Download video info, descriptions, comments, and thumbnails. - **Live Stream Support**: Download live streams. - **AI Video Summary & Processing**: Summarize videos with Gemini, OpenAI, or Ollama, and edit videos with natural language tools. - **Time Range Download**: Download only the segment you need. - **Batch & Playlist**: Download multiple videos or entire playlists. - **Audio Extraction**: Extract audio in MP3, M4A, or Opus formats. - **Subtitle Support**: Download or embed subtitles. - **Subtitle Workshop**: Create, edit, and refine subtitles with timing tools, find

Overview: WorldX is an AI-driven platform that generates a fully autonomous world with AI agents, procedural generation, and emergent narrative. It allows you to describe any scene and watch it evolve into reality. Key features: - **一句话创造世界** - Generate any scene, watch it transform into reality. - **AI生成地图与角色** - Generate maps and characters procedurally, not just templates. - **自主Agent驱动** - AI-driven agents make decisions, interact with the environment, establish relationships, and engage in conversations. - **记忆与人格** - Characters remember past experiences and form unique behaviors. - **多日演化** - Supports continuous evolution across days. - **上帝模式** - Broadcast events, edit character personas/memories, and engage in hypothetical dialogues. - **时间线系统** - Multiple time lines can be created within the same world. - **中英双语** - Support for dual language. Tech stack: TypeScript Why it matters: WorldX is a tool for anyone looking to create an AI-driven, procedurally generated world. It allows for the spontaneous emergence of stories and characters without the need for pre-written scripts, making it a powerful tool for both creators and developers. Repository: YGYOOO/WorldX Primary language: TypeScript GitHub stars: 1576 GitHub description: One sentence creates an AI-driven world — generate maps, characters, and watch stories emerge on their own. 一句话生成一个AI自主驱动的世界. README excerpt: <p align="center"> <img src="docs/logo

Overview: Robotic Process Automation (RPA) framework designed for Python developers, offering a collection of open-source libraries and tools to automate tasks across various applications. Key features: - Robotic Process Automation (RPA) libraries and tools - Integration with Robot Framework - Support for Python development Tech stack: Python Why it matters: This framework provides a comprehensive solution for automating repetitive tasks in applications, enhancing productivity and efficiency in software development.

Overview: This repository contains a curated list of essential papers on textual adversarial attack and defense, focusing on the analysis of adversarial attacks and defenses in natural language processing (NLP). Key features: - RobustQA: A Framework for Adversarial Text Generation Analysis on Question Answering Systems - SeqAttack: On Adversarial Attacks for Named Entity Recognition - OpenAttack: An Open-source Textual Adversarial Attack Toolkit - TextAttack: A Framework for Adversarial Attacks, Data Augmentation, and Adversarial Training in NLP Tech stack: Python Why it matters: This repository serves as a valuable resource for researchers and practitioners in the field of NLP, providing a comprehensive overview of adversarial attacks and defenses, as well as tools for analyzing and mitigating such attacks. Repository: thunlp/TAADpapers Primary language: Python GitHub stars: 1575 GitHub description: Must-read Papers on Textual Adversarial Attack and Defense README excerpt: This list is currently maintained by [Chenghao Yang](https://yangalan123.github.io/) at UChicago. Other previous main contributors include [Fanchao Qi](https://fanchao-qi.github.io/) and [Yuan Zang](https://zangy17.github.io/) when they were at [THUNLP](http://nlp.csai.tsinghua.edu.cn/). We thank all the great contributors very much.

### Overview: This repository contains annotated datasets for named entity recognition (NER) and entity recognition tasks. The datasets cover various languages, domains, and entity types, making it a valuable resource for researchers and developers in the field. ### Key Features: - **Datasets for English NER**: Includes datasets such as CONLL 2003, NIST-IEER, MUC-6, OntoNotes 5, BBN, GMB-1.0.0, GUM-3.1.0, wikigold, Ritter, BTC, WNUT17, i2b2-2006, i2b2-2014, and AnEM. - **Datasets for Other Languages**: Links to NER corpora in other languages, including English, French, German, Spanish, Italian, and more. - **Converting to CoNLL 2003 Format**: Provides code to convert datasets to the CoNLL 2003 format, facilitating interoperability with other tools and models. ### Tech Stack: - Python ### Why It Matters: These annotated datasets are crucial for training and evaluating NER models across different languages, domains, and entity types. They provide a comprehensive set of examples and labels that are essential for improving NER algorithms and ensuring better performance in real-world applications.