Projects

### Overview: AI projects designed for developers to build in a weekend, covering a range of functionalities including personal APIs, pattern mirroring, and more. Each project includes a detailed README with starter code and prompts. ### Key Features: - **Personal APIs**: One local server where your life has addresses. - **Pattern Mirroring**: Years of your camera roll, rendered as the shape of your life. - **Lookout**: Stops you checking pages, watches your sources and speaks only when something changes. - **Past-You Searcher**: Everything you've ever written becomes one memory you can ask questions. - **AI Daily Digest**: 5 sources → one morning email of what matters. - **Auto-Skill Builder**: Corrections in a chat → installable skill file. - **Content Repurposer**: One post → 5 platform-native versions. - **AI Inbox Triage**: Inbox connected → wake to 3 drafts ready. ### Tech Stack: Python ### Why It Matters: Each project is designed to be shippable and ready for developers to implement in their own projects, focusing on functionality rather than workflow hacks.

Overview: This repository implements the Memorizing Transformers model from the ICLR 2022 paper, focusing on an attention net augmented with indexing and retrieval of memories using approximate nearest neighbors. Key features: - Implementation of Memorizing Transformers in Pytorch - Hybrid attention mechanism combining local and distant attention logits - Cosine similarity attention layer for KNN attention - Support for creating and using KNN memories - Clearing memories automatically for batch indices containing specific SOS token IDs Tech stack: Pytorch Why it matters: This implementation provides a practical example of the Memorizing Transformers model, demonstrating how to integrate indexing and retrieval of memories using approximate nearest neighbors, which can be useful in various natural language processing tasks where context is crucial. Repository: lucidrains/memorizing-transformers-pytorch GitHub stars: 646

Overview: Small-Text is an active learning library for text classification, providing state-of-the-art methods for efficient labeling of training data. Key features: - Unified interfaces for active learning, allowing easy mixing of query strategies with classifiers from sklearn, PyTorch, or transformers. - GPU support for PyTorch models, integrating transformers for state-of-the-art text classification. - Pre-implemented query strategies, initialization strategies, and stopping criteria. - Scientifically evaluated components. - Ready to use for biomedical research, social science, information science, computer science, and political communication. Tech stack: Python. Why it matters: Small-Text offers a powerful tool for active learning in text classification, making it an essential resource for researchers and practitioners in various fields. Repository: webis-de/small-text Primary language: Python GitHub stars: 646 GitHub description: Active Learning for Text Classification in Python README excerpt: [](https://pypi.org/project/small-text/) [](https://anaconda.org/conda-forge/small-text) [](https://codecov.io/gh/webis-de/small-text) [![Documentation Status](https://readthedocs

### Overview: UPSNet is a unified panoptic segmentation network that achieves state-of-the-art performance on both COCO and Cityscapes datasets. The model is trained on 16 GPUs and is designed to be easily deployable and adaptable to various hardware configurations. ### Key Features: - Unified architecture for panoptic segmentation - Efficient training on 16 GPUs - Supports both Cityscapes and COCO datasets - Available pretrained weights for reproduction of results ### Tech Stack: - Python - PyTorch - Horovod ### Why It Matters: - Provides a comprehensive solution for panoptic segmentation tasks - Demonstrates the effectiveness of unified architecture in achieving high accuracy - Facilitates easy deployment and scalability across different hardware configurations ### Repository: - **Repository**: `uber-research/UPSNet` - **Stars**: 645 - **Description**: A unified panoptic segmentation network presented in the CVPR 2019 paper. ### Requirements: - **Software**: Anaconda3 - **Hardware**: 4-16 GPUs with at least 11 GB memory ### Installation: - **Clone Repo**: `git clone https://github.com/uber-research/UPSNet` - **Build Modules**: `init.sh` - **Cityscapes Preparation**: `ln -s $CITYSCAPES_ROOT data/cityscapes` and `init_cityscapes.sh` - **COCO Preparation**: `ln -s $COCO_ROOT data/coco` and `init_coco.sh

Overview: Seg-Zero is a research project that demonstrates emergent test-time reasoning ability through a reasoning-chain guided segmentation approach. It achieves superior performance on both in-domain and out-of-domain data compared to supervised fine-tuning. Key features: - Seg-Zero generates a reasoning chain before producing the final segmentation mask. - It is trained exclusively using reinforcement learning, without any explicit supervised reasoning data. - Supports multi-object segmentation and multi-tasks. Tech stack: Python Why it matters: Seg-Zero showcases advanced reasoning capabilities in segmentation tasks, making it a significant contribution to the field of computer vision and reinforcement learning. Repository: JIA-Lab-research/Seg-Zero Primary language: Python GitHub stars: 644 GitHub description: Project Page For "Seg-Zero: Reasoning-Chain Guided Segmentation via Cognitive Reinforcement"

Overview: LarAgent is an AI agent development framework for Laravel, designed to make AI agent creation and maintenance as easy as possible. Key features: - Eloquent-style API for building real AI agents - First-class tooling system with MCP server support - Pluggable memory & context management - Multi-agent workflows with queues, chainable tasks, and reasoning - Structured output for reliability and seamless integration - Multi-modal input and modular provider support - Extensive Event system for customization and observability - Built-in API exposure via OpenAI-compatible schema Tech stack: PHP Why it matters: LarAgent simplifies the process of creating and maintaining AI agents in Laravel projects, offering developer-first tools and a growing ecosystem of Laravel tools powered by Redberry.

### Overview: The Ghost in the Minecraft (GITM) framework integrates Large Language Models (LLMs) with text-based knowledge and memory, aiming to create Generally Capable Agents in Minecraft. GITM features broad task coverage, high success rates, and excellent training efficiency, making it a novel approach to developing capable agents for handling long-horizon, complex tasks in open-world environments. ### Key Features: - **Broad Task Coverage:** GITM can unlock 100% of the Minecraft Overworld technology tree, surpassing previous agents that can only achieve 30% completion. - **High Success Rate:** GITM achieves a 67.5% success rate on the "ObtainDiamond" task, improving upon OpenAI's VPT by +47.5%. - **Excellent Training Efficiency:** GITM requires no GPUs and can be trained in just 2 days using a single CPU node with 32 CPU cores. ### Tech Stack: GITM leverages Large Language Models (LLMs) to create Generally Capable Agents in Minecraft, integrating text-based knowledge and memory to address the challenges of mapping long-horizon and complex goals to sequence of keyboard/mouse operations. ### Why It Matters: This research demonstrates the potential of LLMs in developing capable agents for handling long-horizon, complex tasks and adapting to uncertainties in open-world environments. GITM's ability to unlock 100% of the Minecraft Overworld technology tree and achieve high success rates on tasks like obtaining diamonds and enchanted books highlights its potential for real-world applications in creating autonomous

### Overview: SkyWorkAIGC/SkyPaint-AI-Diffusion is an optimized AI drawing model based on Stable Diffusion, capable of generating high-quality images in various modern art styles using both Chinese and English inputs. ### Key features: - Supports inputting Chinese and English text to generate images. - Generates high-quality images in multiple modern art styles. - Compatible with stable_diffusion_1.x official models and related micro-trained models. ### Tech stack: Python, Stable Diffusion, Hugging Face ### Why it matters: This model leverages the capabilities of Stable Diffusion to create vivid and modern art-inspired images, making it a valuable tool for artists and designers who need to incorporate text into their artwork.

### Overview: Hermes Agent skills and plugins offer a comprehensive toolkit for anyone looking to enhance their AI assistant's capabilities. With over 350 tools, memory providers, and guides, the ecosystem is designed to make your agent smarter and more versatile. ### Key Features: - **Built-in Skills:** 82 skills included in the default installation, ready to use right out of the box. - **Optional Skills Catalog:** 117 additional skills available for customization. - **Community Skills:** 169 skills vetted and curated by the community for quality assurance. - **Memory Providers:** Tools to store and retrieve information for quick access. - **Agent Profiles:** Templates to personalize your agent's behavior. - **Surface Interfaces:** Visual interfaces for interacting with your agent. - **Tools:** Additional utilities for enhancing your agent's functionality. ### Tech Stack: Hermes Agent - **Hermes Agent:** The core agent software from Nous Research. ### Why It Matters: - **Versatility:** The skills in this directory are designed to be install-ready, making it easy to integrate them into your agent's workflow. - **Scalability:** Whether you're running your agent on a $5 VPS, a GPU cluster, or your laptop, the skills are built to scale. - **Learning Loop:** Hermes is the only agent with a real learning loop, making it more powerful and adaptable over time. - **Ease of Use:** The skills are designed to be installable in a matter of minutes, allowing you to start using them immediately. ### Repository: ZeroPointRepo

### Overview: TGO is an open-source AI agent customer service platform designed to help businesses build AI agent teams for effective customer service. It integrates multi-channel access, agent orchestration, knowledge base management, and human collaboration using RAG, LLM orchestration, and human collaboration tools. ### Key Features: - **Multi-Agent Support**: Configure multiple AI agents for different business scenarios. - **Multi-Model Integration**: Connect with various LLM providers (OpenAI, Anthropic, etc.). - **Streaming Response**: Real-time AI responses via SSE for smooth conversation experience. - **Context Memory**: Maintain conversation history for coherent dialogue. ### Tech Stack: TypeScript ### Why It Matters: TGO simplifies the process of building AI agent teams, making it easier for businesses to provide efficient and personalized customer service. Its comprehensive features and user-friendly interface make it an ideal solution for modern customer service needs. ### Repository Structure: - **tgo-ai**: AI/ML operations service for managing agents, tool bindings, knowledge bases, and usage analytics. - **tgo-api**: Core business logic service handling user management, visitor tracking, assignment, and communication. - **tgo-cli**: CLI tool & MCP Server enabling AI agents to execute customer service operations.

### Overview: OpenHowNet is a Python-based API designed to facilitate the search, display, and analysis of HowNet and OpenHowNet data. This project serves as a comprehensive resource for researchers and developers in natural language processing, providing access to sememe-based lexical knowledge and various linguistic annotations. ### Key Features: - **HowNet Dictionary**: A core data structure that encapsulates the HowNet dictionary, including definitions, POS tags, sentiment orientations, and example sentences for both Chinese and English words. - **Sense Class**: Manages the information of HowNet concepts, including Chinese and English words, POS tags, sememe-based definitions, sentiment orientations, and example sentences. - **API Functions**: Offers functionalities for retrieving HowNet data, displaying sememe trees, calculating word similarity based on sememes, and more. ### Tech Stack: - Python 3.6+ - anytree>=2.4.3 - tqdm>=4.31.1 - requests>=2.22.0 ### Why It Matters: OpenHowNet provides a reliable and efficient way to access and analyze sememe-based lexical knowledge, making it invaluable for researchers and developers in natural language processing. Its API simplifies the process of querying and utilizing HowNet data, enhancing the capabilities of NLP applications.

### Overview VirtualHome is an interactive platform that simulates complex household activities through a Python API. Key features include procedural generation capabilities, customizable environments, and support for multi-agent interactions. The platform allows for the rendering of activities in a virtual environment, enabling the simulation of various scenarios and interactions with the environment. ### Key Features - Procedural Generation Capabilities - Customized Environments - Multi-Agent Simulations - Realistic Physics and Lighting - Time Management System - Streamable Ground-Truth Data ### Tech Stack Python ### Why It Matters VirtualHome provides a versatile tool for researchers and developers interested in embodied AI and multi-agent simulations. It offers a flexible environment for training and testing agents, making it useful for various research and educational purposes. ### Repository xavierpuigf/virtualhome ### GitHub Stars 643 ### Primary Language Python ### GitHub Description API to run VirtualHome, a Multi-Agent Household Simulator ### Contents 1. Overview 2. Set Up 3. Generating Videos/Keyframes 4. VirtualHome RL Env 5. Dataset 6. Modify VirtualHome 7. Citation 8. Contributors

### Overview: calm is an implementation of language model inference that aims to maximize single-GPU single-batch hardware utilization for large language model architectures. This project is focused on experimentation and prototyping, with minimal dependencies and no intention of being production-ready. ### Key Features: - CUDA/Metal-accelerated language model inference - Supports a subset of decoder-only transformer architectures - Tested on a variety of models including Llama-like, Yi, Mistral, Qwen2, and Mixtral ### Tech Stack: C ### Why it Matters: This project provides a lightweight and efficient way to run language models, leveraging CUDA and Metal for acceleration, making it suitable for experiments and prototyping without the need for complex dependencies.

Overview: OfficeIMO is a collection of .NET libraries that can convert, read, edit, and export various Office and document formats, including OneNote, PDF, Markdown, RTF, OpenDocument, and more, without requiring Microsoft Office or its automation. It is designed to run in various environments, including services, desktop applications, build agents, and containers, and is MIT-licensed. Key features: - Converts and exports Office and document formats - Runs in various environments without requiring Microsoft Office - Supports OneNote, PDF, Markdown, RTF, OpenDocument, AsciiDoc, LaTeX, OPML, DocBook, CSV, EPUB, ZIP, drawing, Apple iWork source, legacy Word `.doc`, legacy Excel `.xls`, and legacy PowerPoint `.ppt`/`.pot`/`.pps` - Composes package-owned document models and exposes result-bearing APIs for fidelity diagnostics Tech stack: C# Why it matters: OfficeIMO is a powerful tool for developers looking to integrate Office and document formats into their applications without the need for Microsoft Office, making it ideal for automation, services, and legacy applications.

Overview: Carrel is a self-hosted knowledge base designed for AI agents, automating document parsing, indexing, and retrieval. It integrates with agents like Claude Code and Codex, providing access to various files and documents, and supports knowledge graphs for enhanced search capabilities. Key features: - Document parsing: Supports PDF, Word, PowerPoint, spreadsheets, images, HTML, Markdown, notebooks, source code, and workflow files. - Automatic updates: Scans for file changes, updates indexes, and maintains enabled knowledge graphs through scheduled tasks. - Hybrid retrieval: Combines vector, keyword, graph, and visual search, with automatic knowledge-base selection and optional text/image reranking. - Knowledge graphs: Extracts entity types, relationships, and facts from documents; retains sources, time information, and conflict indicators. - Agent access: Searches through an API, fetches surrounding text or original images, and follows graph relationships when more evidence is needed. - Web console: Configures processing, inspects chunks and graphs, tracks tasks, and manages services. Tech stack: Python Why it matters: Carrel simplifies the process of accessing and integrating various documents and files into AI agents, enhancing their ability to retrieve and analyze information effectively. Repository: FengZhenfei/carrel Primary language: Python GitHub stars: 643 GitHub description: Ontology-Augmented Generation for AI agents.

### Overview: ANOLISA is a multi-component system software stack designed for AI Agent workloads, built on existing operating system kernels and infrastructure. It provides context management, Skills delivery, workspace recovery, security, and observability for Agent applications, frameworks, and platforms. ### Key features: - **Context Management**: Ensures seamless integration with existing systems. - **Skills Delivery**: Optimized execution of AI Agent tasks. - **Workspace Recovery**: Safely restores Agent workspaces. - **Security**: Robust protection against threats. - **Observability**: Detailed tracking of Agent activities. ### Tech stack: Rust ### Why it matters: ANOLISA offers a robust, secure, and efficient solution for AI Agent workloads, making it an essential tool for developers and system administrators alike. Its open architecture allows for easy integration with existing systems and platforms, making it a versatile choice for various use cases.

Overview: BiliSum is an AI-powered video summarizer and knowledge base designed for Bilibili, YouTube, and local videos. It automatically transcribes, summarizes, and generates notes, mind maps, and knowledge bases, all stored locally. Key features: - Transcription - Text notes - Image notes - Mind maps - Knowledge base Tech stack: Python 3.12 Why it matters: BiliSum enhances the Bilibili experience by summarizing videos, providing text and image notes, generating mind maps, and creating a knowledge base. It also supports local video summarization and transcriptions, making it versatile for various video platforms. Repository: lycohana/BiliSum GitHub stars: 646 GitHub description: AI video summarizer and knowledge base for Bilibili, YouTube, and local videos. Automatically transcribes, summarizes, and generates notes, mind maps, and knowledge bases. All data is stored locally.

Overview: This project is an implementation of Sheared LLaMA, a method for accelerating language model pre-training by structurally pruning strong base models. The code is based on Composer, a package designed for large language model pre-training, and includes pruning logic and sample scripts for both pruning and continued pre-training. Key features: - Pruning strong base models to obtain strong small-scale language models - Pruning LLaMA-2.7B to produce a model as strong as an OpenLLaMA model with 3% of its pre-training cost - Instruction-tuned models for better performance - Pruning and continued pre-training scripts for various models Tech stack: Python Why it matters: This project provides a cost-effective way to accelerate language model pre-training by structurally pruning strong base models, making it an important contribution to the field of large language model pre-training.

Overview: EasyRAG is an efficient retrieval-augmented generation framework designed for automated network operations. It achieved top positions in the GLM4 track of the CCF AIOps International Challenge 2024. Key features: - Accurate question answering through a straightforward RAG scheme - Simple deployment with minimal VRAM requirements - Efficient inference for improved system efficiency Tech stack: Python Why it matters: EasyRAG stands out with its accuracy, simplicity, and efficiency, making it a valuable tool for automated network operations. Repository: BUAADreamer/EasyRAG GitHub stars: 641

### Overview: TextCNN Pytorch实现 中文文本分类，情感分析。使用PyTorch实现，论文来自[Convolutional Neural Networks for Sentence Classification](https://arxiv.org/abs/1408.5882)。 ### Key features: - 使用PyTorch实现文本分类。 - 支持中文文本分类和情感分析。 - 提供多种训练方法，包括随机初始化、静态词向量、微调预训练词向量和微调加静态词向量。 ### Tech stack: Python ### Why it matters: 该实现展示了如何使用TextCNN模型进行中文文本分类和情感分析，提供了多种训练方法和词向量微调策略，有助于理解和优化文本分类任务。