Projects

### Overview: Audio diffusion using PyTorch is a comprehensive library designed for generating audio from scratch using diffusion models. This project leverages PyTorch to create a versatile audio generation system, including unconditional and text-conditioned models, and various diffusion methods for upsampling and vocoding. ### Key Features: - **Unconditional Audio Generation**: Generates audio from scratch without any external conditioning. - **Text-Conditioned Audio Generation**: Uses T5-base embeddings to condition the audio generation process. - **Diffusion Autoencoding**: Applies diffusion to autoencode audio, improving quality. - **Upsampling**: Converts audio from a lower sample rate to a higher one, such as from 3kHz to 48kHz. - **Vocoding**: Enhances audio quality through vocoding techniques. ### Tech Stack: - **PyTorch**: The primary framework used for implementing the audio diffusion models and diffusion methods. ### Why It Matters: This library provides a powerful toolset for audio generation and manipulation, making it valuable for developers and researchers looking to create high-quality audio content without the need for pre-trained models. The versatility of the models and the ability to customize diffusion methods make it suitable for various applications, including music synthesis, sound design, and more. ### Installation: ```bash pip install audio-diffusion-pytorch ``` ### Usage: - **Unconditional Generator**: Generates audio from scratch. - **Text-Conditional Generator**: Conditions the audio generation with text embeddings. - **Diffusion Upsampler**: Converts audio from a lower sample rate to a higher one

### Overview: In-Context Edit: A Novel Image Editing Approach In-Context Edit is a groundbreaking image editing method that achieves state-of-the-art results using just 0.5% of the training data and 1% of the parameters required by prior SOTA methods. This approach enables high-precision multi-turn image editing with visually impressive single-turn results. ### Key Features: - **In-Context Generation**: Enables precise image editing with high precision. - **State-of-the-Art Results**: Surpasses GPT-4 in ID persistence. - **MoE ckpt**: Only 4GB VRAM is sufficient to run. ### Tech Stack: Python ### Why It Matters: In-Context Edit is a novel approach that achieves state-of-the-art results using just 0.5% of the training data and 1% of the parameters required by prior SOTA methods. This method enables high-precision multi-turn image editing with visually impressive single-turn results. Repository: [River-Zhang/ICEdit](https://github.com/River-Zhang/ICEdit) Primary Language: Python GitHub Stars: 2098

### Overview: Gini Agent is a personal agent that remembers, improves, and runs without forcing you to read a log line. ### Key features: - Persistent chat, runs, tasks, approvals, traces, audit events, jobs, memories, and skills. - Approval-gated file, terminal, and code tools. - Support for multiple providers: Codex OAuth; OpenAI / Azure OpenAI / DeepSeek / OpenRouter API keys; Anthropic Claude API; Amazon Bedrock. ### Tech stack: TypeScript. ### Why it matters: Gini Agent is designed to be a product, not plumbing, making it accessible and secure. It integrates seamlessly across devices, providing a secure and efficient way to manage tasks and interactions.

### Overview: .NET for Apache Spark™ simplifies the integration of Apache Spark™ into .NET applications, enabling developers to leverage the powerful data processing capabilities of Apache Spark™ within their C# and F# projects. ### Key Features: - High-performance APIs for working with structured data and processing streaming data. - Compatible with Apache Spark™ 3.5 through 3.5.4, targeting .NET Standard 2.0 and 2.1. - Supports .NET 8 and .NET Framework 4.8. - Provides a seamless experience for developers familiar with .NET, leveraging the .NET ecosystem. ### Tech Stack: .NET for Apache Spark™ leverages the .NET Framework and .NET 8, targeting .NET Standard 2.0 and 2.1. ### Why It Matters: .NET for Apache Spark™ streamlines the integration of Apache Spark™ into .NET applications, making it easier for developers to leverage the capabilities of Apache Spark™ within their .NET projects. This tool simplifies the process of accessing and processing large-scale data, enhancing the performance and efficiency of data-driven applications.

Overview: The AI-Aimbot project is a collection of scripts designed to simulate AI-driven aimbots for various first-person shooter games, including CS2, Valorant, Fortnite, and Apex Legends. It leverages the YOLOv5 model for its functionality. Key features: - Fast 🏃‍♂️: A basic version that can be set up on any computer. - Faster 🏃‍♂️💨: An optimized version that requires editing a file and works on any computer. - Fastest 🚀: An enterprise-level version that requires a specific hardware configuration and works only on computers with Nvidia GPUs. Tech stack: The project uses Python, PyTorch, and YOLOv5. Why it matters: The AI-Aimbot serves as an educational tool to highlight the vulnerabilities of game developers to AI-driven cheats, encouraging developers to improve their security measures.

### Overview: Magicoder is a model family that leverages OSS-Instruct to generate low-bias and high-quality instruction data for code. This approach mitigates inherent biases in LLM-synthesized instruction data by providing a wealth of open-source references. ### Key Features: - **Magicoder-CL-7B**: 7B size, 60.4 (+55.5) on HumanEval - **Magicoder-*S*-CL-7B**: 7B size, 70.7 (+66.5) on HumanEval - **Magicoder-DS-6.7B**: 6.7B size, 66.5 (+60.4) on HumanEval - **Magicoder-*S*-DS-6.7B**: 6.7B size, **76.8** (+**70.7**) on HumanEval ### Tech Stack: Python ### Why It Matters: Magicoder and OSS-Instruct have inspired numerous projects, including Google CodeGemma, OpenCodeInterpreter, RTLCoder, CodeUltraFeedback, and speechless-coder-ds-6.7b. It has also demonstrated superior performance in HumanEval, outperforming GPT-3.5-turbo-1106 and Gemini Ultra by 76.8% (+70.7%). ### Repository: - Repository: ise-uiuc/magicoder - GitHub Stars: 2094 - GitHub Description: [ICML'24] Magic

Overview: 本书是李航老师的《统计学习方法》和《机器学习方法》的习题解答，涵盖了监督学习、无监督学习和深度学习的主要方法。读者可以通过详细的解答和示例代码，深入理解机器学习算法原理。 Key features: - 监督学习：感知机、k近邻法、朴素贝叶斯法、决策树、逻辑斯谛回归与最大熵模型、支持向量机、提升方法、EM算法、隐马尔可夫模型和条件随机场。 - 无监督学习：聚类方法、奇异值分解、主成分分析、潜在语义分析、概率潜在语义分析、马尔科夫链蒙特卡罗法、潜在狄利克雷分配和PageRank算法。 - 深度学习：前馈神经网络、卷积神经网络、循环神经网络、序列到序列模型、预训练语言模型和生成对抗网络。 Tech stack: Jupyter Notebook Why it matters: 本书为机器学习初学者和算法工程师提供了一个系统的学习资源，帮助他们理解和实现机器学习算法。通过习题解答和示例代码，读者可以巩固理论知识并提升实践技能。

### Overview: ChatWiki is a workflow automation platform designed for the WeChat ecosystem, enabling every official account to become a super AI agent. It seamlessly integrates with the official account platform, allowing for easy creation of WeChat ecosystem applications through drag-and-drop, and supports features like one-click rewriting of official account articles and AI-curated comment replies. ### Key Features: - **Deep Integration with WeChat Ecosystem**: Supports automatic replies to private messages for unverified official accounts, including text, voice, images, mini-program cards, and video messages. - **WeChat Workflows**: Integrates trigger scenarios such as user private messages, comments, follows, unfollows, menu clicks, etc., and supports various processing steps like replying to private messages, tagging fans, generating draft articles, publishing articles, and more. - **Knowledge Base Synchronization**: Scrapes articles and materials from official accounts to build a knowledge base with a single click. - **Basic Capabilities**: Conversational workflows, plugin workflows, including basic workflow nodes, bidirectional MCP, Agent mode, and user interaction. - **Document Knowledge Base**: Supports URL reading, batch document import, API integration, AI-based segmentation, QA segmentation, parent-child segmentation, and knowledge graphs, hybrid vector search, and visual exploration of knowledge graphs. - **QA Knowledge Base**: Automatically extracts QA knowledge from uploaded documents, supports automatic clustering of unknown questions, and summarizes common FAQs from human conversations. - **Human Handoff**: Handles general user inquiries via bot, while also supporting human staff. Issues that the bot cannot resolve can be escalated to

Overview: ELF is an extensive, lightweight, and flexible platform for game research, particularly designed for real-time strategy (RTS) games. It provides miniature game environments, concurrent simulation, intuitive APIs, web-based visualization, and a reinforcement learning backend powered by PyTorch. Key features: - End-to-End: Offers miniature real-time strategy game environments and concurrent simulation. - Extensive: Can plug any game with a C/C++ interface into the framework. - Lightweight: Runs fast with minimal overhead, achieving 40K frames per second per core on a MacBook Pro. - Flexible: Flexible pairing between environments and actors, supporting single agent, multiple agents, and multiple environments with one actor. Tech stack: C++ and Python. Why it matters: ELF provides a comprehensive solution for game research, offering real-time strategy games with high simulation speed, efficient resource usage, and a powerful reinforcement learning backend. This makes it suitable for various research and development projects in the field of game research and reinforcement learning. Repository: facebookresearch/ELF GitHub stars: 2090 GitHub description: An End-To-End, Lightweight and Flexible Platform for Game Research

### Overview: This repository is a curated list of resources related to the trustworthiness of large models across multiple dimensions, focusing on multi-modal models such as vision-language models and diffusion models. The list includes papers, benchmarks, and other resources that help ensure the safety, security, and privacy of these models. ### Key features: - **Papers**: A collection of papers related to the safety, security, and privacy of large language models. - **Benchmarks**: Various benchmarks and metrics used to evaluate the performance of these models. - **Comments**: A section highlighting new datasets, agents, code, venues, and more. - **Venues**: Information about the conferences and blogs where papers have been accepted. - **Updates**: A call to action for contributors to help update the list. ### Tech stack: Unknown ### Why it matters: This repository serves as a valuable resource for researchers, developers, and practitioners interested in ensuring the safety, security, and privacy of large language models. By providing a curated list of relevant papers, benchmarks, and other resources, it helps in maintaining a high standard of trustworthiness in these models across various dimensions.

### Overview: This repository contains Russian GPT-3 models trained on a large dataset of Russian language. The models include ruGPT3XL, ruGPT3Large, ruGPT3Medium, ruGPT3Small, and ruGPT2Large. These models are autoregressive transformer language models with 2048 and 1024 sequence lengths, respectively, and support sparse and dense attention blocks. The repository provides detailed usage examples and fine-tuning instructions. ### Key features: - Russian GPT-3 models (ruGPT3XL, ruGPT3Large, ruGPT3Medium, ruGPT3Small, ruGPT2Large) - Sparse and dense attention blocks - Fine-tuning instructions for ruGPT3 models - Usage examples in Colab notebooks ### Tech stack: Python ### Why it matters: These models are useful for various applications involving Russian language processing, such as text generation, fine-tuning, and other NLP tasks. They provide a powerful toolset for developers working with Russian language data. ### GitHub stars: 2089 ### Repository: * ai-forever/ru-gpts ### Primary language: Python ### OpenSource Solutions with ruGPT3: - Fine-tuning scripts provided - Colab notebooks for model generation and fine-tuning ### Papers mentioning ruGPT3: - Papers discussing the use and performance of the ruGPT3 models ### Table of contents: - ruGPT3XL - Setup -

Overview: This project is a web-based interface for running Large Language Models (LLMs) using Colab and various pretrained models. Key features: - Colab-based web interface - Support for multiple models including Vicuna, Stable-Vicuna, GPT4-X, Pygmalion, Koala, OASST-Llama, WizardLM, MPT-Storywriter, and Falcon - Integration with Twitter, Discord, and Patreon communities for updates and engagement - Colab notebooks for detailed model information and model pages Tech stack: Python, Jupyter Notebook, Colab, Hugging Face Transformers, and various pretrained models. Why it matters: This project provides a user-friendly interface for experimenting with different LLMs, making it accessible for developers and researchers alike.

Overview: OnnxStream is a lightweight inference library designed for running models like Stable Diffusion 1.5 on devices with limited RAM, such as a Raspberry Pi Zero 2. It achieves this by decoupling the inference engine from the model weights, allowing for efficient memory usage and minimal RAM footprint. Key features: - Supports Stable Diffusion 1.5, 1.0 Base, and 1.0 Turbo models. - Runs Stable Diffusion XL 1.0 on a RPI Zero 2 (298MB RAM). - Supports Mistral 7B on desktops and servers. - ARM, x86, WASM, and RISC-V architectures supported. - Accelerated by XNNPACK. - Python, C#, and JS(WASM) bindings available. Tech stack: C++. Why it matters: OnnxStream is crucial for developers looking to run large models on constrained devices, ensuring efficient memory usage and minimal resource consumption without compromising performance. Repository: vitoplantamura/OnnxStream GitHub stars: 2088 GitHub description: Lightweight inference library for ONNX files, written in C++. It can run Stable Diffusion XL 1.0 on a RPI Zero 2 (or in 298MB of RAM) but also Mistral 7B on desktops and servers. ARM, x86, WASM, RISC-V supported. Accelerated by XNNPACK. Python, C# and JS(WASM) bindings available. Repository: vitoplantamura/On

Overview: Ui.Vision is a powerful browser and desktop automation tool that combines browser automation, OCR, and image recognition with an MCP server for AI assistants. It allows users to record macros, write JavaScript, or ask AI assistants to author macros, making it ideal for automating websites, extracting data, and controlling desktop applications. Key features: - Record macros or write JavaScript to automate websites and desktop apps. - Use OCR and image recognition for visual tasks. - Connect to an MCP server for AI-assisted automation. - Save and run macros, and inspect logs and screenshots. Tech stack: TypeScript. Why it matters: Ui.Vision simplifies the process of automating tasks across different platforms, making it a valuable tool for developers and businesses looking to streamline their workflows and improve efficiency.

### Overview: docext is an on-premises document intelligence toolkit powered by vision-language models (VLMs). It offers three core capabilities: PDF and image to markdown conversion, document information extraction, and an intelligent document processing benchmarking platform. ### Key Features: - **PDF and Image to Markdown Conversion**: Convert both PDF and images to markdown with intelligent content recognition and semantic tagging. - **LaTeX Equation Recognition**: Convert inline and block LaTeX equations in images to markdown. - **Intelligent Image Description**: Generate detailed descriptions for all images in the document. - **Signature Detection**: Detect and mark signatures and watermarks. - **Watermark Detection**: Detect and mark watermarks. - **Page Number Detection**: Detect and mark page numbers. - **Form Elements**: Convert form checkboxes and radio buttons into standardized Unicode symbols. - **Table Detection**: Convert complex tables into HTML tables. ### Tech Stack: Python ### Why It Matters: docext is a comprehensive toolkit for document intelligence, offering robust OCR-free conversion and structured information extraction from various document types. The intelligent document processing benchmarking platform provides a valuable resource for evaluating and benchmarking vision-language model performance across key document intelligence tasks.

Overview: Waku Agent is a local-first AI assistant that combines the four pillars of a serious agent: Harness, Loop, Memory, and Eval/LLM-ops. It is built with Python and designed to be legible as it grows. Key features: - Local-first: Your memory is stored in a SQLite file, which you can open, read, and modify. - Memory is the hero: It uses semantic, episodic, and procedural memory, with a gate for deciding what to remember and a pass for what to keep. - The loop is 95 lines of Python code, making it easy to step through. - A local dashboard displays messages as they flow through the harness, allowing you to watch how your AI thinks. - Eval built in: It includes deterministic tests and LLM-as-judge, side by side, with a release gate. Tech stack: Python Why it matters: Waku Agent is designed to be legible as it grows, making it accessible and maintainable for developers. It combines the best practices of serious AI assistants, making it a practical solution for those looking to create their own AI agents. Repository: ShenSeanChen/waku-agent GitHub stars: 2083 GitHub description: Waku Waku! Waku Agent is a local-first AI assistant that shows the four pillars behind every serious agent: Harness · Loop · Memory · Eval/LLM-Ops. No frameworks hiding the good parts. Built by [seanchen.io](https://seanchen.io). Primary language: Python GitHub description:

Overview: PyPOTS is a comprehensive Python toolkit designed for machine learning on partially-observed time series, offering state-of-the-art neural network models for tasks such as imputation, classification, clustering, forecasting, anomaly detection, and cleaning. This toolkit is particularly useful for scientific analysis of industrial irregularly-sampled multivariate time series with missing values. Key features: - State-of-the-art neural network models - Support for imputation, classification, clustering, forecasting, anomaly detection, and cleaning - Extensive support for incomplete industrial irregularly-sampled multivariate time series with NaN missing values - Compatible with various data processing libraries like Pandas and NumPy Tech stack: PyTorch and Python Why it matters: PyPOTS simplifies the process of applying machine learning techniques to time series data, making it easier to perform scientific analysis on complex, incomplete datasets. Its advanced models and robust feature set make it a valuable tool for researchers and practitioners in various fields, including finance, engineering, and environmental science.

### Overview: Parlor is a fully on-device, real-time multimodal AI project that aims to replicate the features of GPT-Live, offering a voice assistant experience similar to ChatGPT Advanced Voice. The software is designed to run on a MacBook M3 Pro and is currently in development. ### Key features: - **Real-time Multimodal AI**: Parlor is built to provide a voice assistant experience that is similar to GPT-Live, using both text and speech. - **On-Device Execution**: The project is designed to run entirely on-device, ensuring privacy and security. - **Fine-Tuning**: The project uses a classic cascade system, with Gemma 4 E4B via llama.cpp for speech synthesis, and Kokoro TTS for speech synthesis. - **WebSocket Communication**: The project uses WebSocket for real-time communication between the browser and the server, allowing for seamless interaction. - **Hands-Free Turn-Taking**: The project includes a hands-free turn-taking system that ensures the conversation remains focused on the user's input. ### Tech stack: Python ### Why it matters: Parlor aims to provide a voice assistant experience that is similar to GPT-Live, offering a more personalized and efficient user experience. The project is designed to be fully on-device, ensuring privacy and security, and is currently in development. ### Repository: fikrikarim/parlor ### Primary language: Python ### GitHub stars: 2076 ### GitHub description: On-device, real-time multimodal AI with features similar to GPT-Live.

Overview: 这个项目是一个专注于推荐系统、计算广告和机器学习领域的技术资料仓库，涵盖了从基础教程到前沿比赛的全面内容。 Key features: - 推荐系统简明教程 - 计算广告系列视屏和讲义 - 逻辑回归模型简介 - FFM讲解PPT - FFM原理与实践 - GBDT算法原理与系统设计简介 - LightGBM、XGBoost、LIBFFM、xLearn、DeepCTR实践代码 - CriteoDisplayAdvertisingChallenge、AvazuClickThroughRatePrediction、2018 IJCAI 阿里妈妈搜索广告转化预测、2018腾讯广告算法大赛相关技术博客和实践代码 Tech stack: Python, R, Java, C++等。 Why it matters: 该仓库提供了从基础学习到实际应用的全面指导，帮助开发者在推荐系统、计算广告和机器学习领域不断提升技能。

Overview: Orbit is a Python package designed for Bayesian forecasting tasks, offering an intuitive interface for time series analysis. It utilizes probabilistic programming languages under the hood, supporting models like Exponential Smoothing, Local Global Trend, Damped Local Trend, and Kernel Time-based Regression. Key features: - Bayesian forecasting with object-oriented design - Probabilistic models for accurate predictions - Support for various sampling and optimization methods Tech stack: Python Why it matters: Orbit simplifies Bayesian forecasting by providing a user-friendly interface while leveraging advanced probabilistic programming techniques, making it valuable for researchers and practitioners in the field of time series analysis and machine learning. Repository: The repository is hosted on GitHub at [https://github.com/uber/orbit](https://github.com/uber/orbit). GitHub stars: 2083 GitHub description: A Python package for Bayesian forecasting with object-oriented design and probabilistic models under the hood.  ------------------------------------------------------------------------------------------------------------------------------------- <!--- BADGES: START --->  [][#pypi-package] [](https://github.com/uber/orbit/actions/workflows/test.yaml) [![Documentation Status](https://readthedocs.org