Projects

### Overview: MemeMeow 是一个基于自然语言的表情包检索工具。它通过自动标注表情包包含义，并通过描述场景或心情快速找到合适的表情包。无需记住具体的文件名或标签，只需输入描述性文字即可。 ### Key features: - **🧠 自动化语境理解**: 使用 OpenCode Agent 分析上传图片，生成结构化分析、语境和描述。 - **📤 直接上传**: 上传图片后，生成文本和视觉向量，后台自动命名并更新文本向量。 - **🤖 自然语言检索**: 采用 Embedding 模型，实现 Q&A 式检索，根据描述找到匹配的图片。 - **👁️ 视觉近邻搜索**: 支持 DINOv2 视觉模型，实现图片之间的相似度搜索。 - **📦 容器部署**: 提供 Vue 3 Web 界面和统一的 API，服务可通过 Docker Compose 启动。 ### Tech stack: Python ### Why it matters: MemeMeow 通过自动化处理和检索功能，大大简化了表情包的管理和使用过程，提高了工作效率和用户体验。

### Overview: This repository contains a collection of data science and machine learning projects with source code. It includes articles, GitHub repositories, and Kaggle kernels that provide practical examples and projects for data science enthusiasts and learners. ### Key Features: - **Data Science Projects:** - President Heights Data Science Project - Birth Rate Analysis Data Science Project - Time Series Data Science Project - Area and Population Data Science Project - Sentiment Analysis Project in R - Detecting Fake News with Python - Detecting Parkinson’s Disease - Breast Cancer Classification - Customer Segmentation using Machine Learning in R - Movie Recommendation System Project in R - Detect Credit Card Fraud with Machine Learning in R - Uber Data Analysis Project - Handwritten Digit Recognition using Python - Gender and Age Detection with OpenCV - Speech Emotion Recognition - Credit Card Fraud Detection Project - Home Loan prediction - Spam Classification (NLP) - Drug Classification using state-of-the-art ML Algo - Box Office Revenue Analysis and Visualization - Heart Disease - Classifications (Machine Learning) - 🧉Birdcall Recognition: EDA and Audio FE - Coronavirus (COVID-19) Visualization & Prediction - Santander EDA and Prediction ### Tech Stack: - Python - R - Kaggle - OpenCV - TensorFlow - Scikit-learn - Keras - PyTorch - Numpy - Pandas

Overview: GNES is a cloud-native semantic search system based on deep neural network, enabling large-scale index and semantic search for text-to-text, image-to-image, video-to-video, and any-to-any content forms. Key features: - Cloud-native and elastic architecture - Easy-to-use interface - State-of-the-art performance leveraging best-of-breed deep learning models Tech stack: Python Why it matters: GNES provides a powerful and flexible solution for semantic search across various content forms, making it suitable for developers and researchers looking to integrate AI-powered search capabilities into their projects.

Overview: Heym is a self-hosted runtime and platform designed for building, orchestrating, running, observing, evaluating, and controlling agentic systems on your infrastructure. It allows you to visually compose agents and workflows, generate them with AI, connect models and systems, and operate them with approval checkpoints, retries, and visibility into execution and model costs. Key features: - Visual workflow composition - AI-generated workflows - Model and system connections - Approval checkpoints - Retries - Execution history tracking - Human review and control Tech stack: Python Why it matters: Heym provides a robust and self-hosted solution for building, operating, and monitoring agentic systems, ensuring confidence in the execution of complex processes and workflows.

### Overview: 本项目旨在为初学者提供简化版本的简单CV项目，包括注意力、分类、检测、关键点检测等。项目内容来自GitHub，不用于商业用途。欢迎在Issues中提出和补充建议。 ### Key features: - **Attention Module** - **Plug-and-play Module** - **Vision Transformer** ### Tech stack: Jupyter Notebook ### Why it matters: 简化版本的模型文件，便于初学者学习和理解计算机视觉领域的知识。

### Overview: CoALA is a comprehensive collection of language agents that utilize the Cognitive Architectures for Language Agents (CoALA) framework. This framework provides a structured approach to building language agents, focusing on both reasoning and grounding actions. The agents are designed to interact with both external and internal environments, utilizing short-term working memory and long-term memories for various tasks. ### Key Features: - **Action Space**: CoALA specifies a language agent's action space, which includes both external and internal actions. The external actions are for interacting with the environment, while the internal actions are for reasoning, retrieval, and learning. - **Decision Making**: The agent's actions are structured into decision-making cycles, with two stages: planning and execution. The planning stage involves reasoning and retrieval actions, while the execution stage selects a learning or grounding action. - **Citations**: The repository includes a BibTeX file with 300+ related citations, including the CoALA paper and other works cited in the paper. ### Tech Stack: - **Programming Language**: TeX ### Why It Matters: CoALA provides a structured and comprehensive framework for building language agents, making it a valuable resource for researchers and developers interested in creating intelligent language agents. The framework's focus on reasoning and grounding actions makes it particularly useful for agents that need to interact with both the external world and their internal memories. ### Repository: Repository: ysymyth/awesome-language-agents GitHub Stars: 1261 GitHub Description: List of language agents based on the "Cognitive Architectures for Language Agents" paper.

Overview: OpenPifPaf is an open-source implementation of the "OpenPifPaf: Composite Fields for Semantic Keypoint Detection and Spatio-Temporal Association" paper, focusing on real-time pose detection and tracking. It is designed to be used with PyTorch and includes a generic neural network architecture and Temporal Composite Association Field (TCAF) for temporal associations. Key features: - Real-time pose detection and tracking - Joint detection and spatio-temporal keypoint association - Composed fields for semantic keypoints - Temporal composite association field (TCAF) for temporal associations Tech stack: Python Why it matters: OpenPifPaf is a significant contribution to the field of computer vision, providing a real-time solution for pose detection and tracking, which is essential for applications like self-driving cars, robotics, and more. Its ability to be deployed on various operating systems and its competitive performance make it a valuable tool for developers working on similar projects.

Overview: GardenPulse is a method-agnostic smart garden companion designed for urban environments, adapting to various growing methods like soil, hydroponics, containers, and indoor setups. It integrates location-aware intelligence, privacy-first AI diagnostics, and seamless weather integration. Key features: - **Method-Agnostic Plant ID**: Identifies plant species and growth method. - **Localized AI Agronomist**: Evaluates leaf patterns on-device. - **Smart Weather Integration**: Adjusts care schedules based on live weather forecasts. - **Gemini Vision Diagnostics**: Runs all leaf-scan diagnostics on-device. Tech stack: TypeScript, React Native, Expo, Firebase, Gemini AI, and Privacy-first architecture. Why it matters: GardenPulse simplifies gardening by providing personalized, method-specific advice tailored to your growing environment, local climate, and live weather conditions.

Overview: This repository is an out-of-box human parsing representation extractor designed for various downstream applications. It includes pretrained models on three popular datasets: LIP, ATR, and Pascal-Person-Part, along with training and inference code. Key features: - Out-of-box human parsing extractor - Pretrained models on three popular single person human parsing datasets - Training and inference code - Simple yet effective extension for multi-person and video human parsing tasks Tech stack: Jupyter Notebook Why it matters: The repository's models rank 1st for all human parsing tracks in the third LIP challenge, demonstrating its effectiveness and reliability in various applications. Repository: GoGoDuck912/Self-Correction-Human-Parsing Primary language: Jupyter Notebook GitHub stars: 1259 GitHub description: An out-of-box human parsing representation extractor.

### Overview: DragDiffusion is an official code for the CVPR 2024 paper "Harnessing Diffusion Models for Interactive Point-based Image Editing". This project aims to leverage diffusion models for interactive point-based image editing, providing users with a powerful tool for creating high-quality images. ### Key Features: - **Interactive Point-based Image Editing**: Users can directly interact with specific points in an image to modify it. - **Support for Diffusers**: The project supports the latest diffusers, ensuring compatibility with the latest advancements in the field. - **Training with LoRA**: Users can train a LoRA (Low-Rank Auto-Encoder) to enhance the quality of their edits. ### Tech Stack: The project is built using Python, leveraging libraries such as PyTorch, OpenCV, and others. ### Why It Matters: DragDiffusion offers a unique and powerful way to interact with images, making it a valuable tool for artists, designers, and anyone looking to experiment with point-based image editing. Its integration with diffusion models ensures that the edits are not only interactive but also highly effective and versatile. ### Repository: The repository is hosted on GitHub under the name Yujun-Shi/DragDiffusion. It has received 1,259 GitHub stars, indicating a high level of interest and engagement from the community. ### Installation: To run DragDiffusion, users need to have a Nvidia GPU with a Linux system and around 14 GB of GPU memory. The project is currently in development, and users are encouraged to report any issues or suggest improvements. ###

### Overview: This survey provides a comprehensive review of efficient large language models, focusing on three main categories: model-centric, data-centric, and framework-centric perspectives. The survey aims to help researchers and practitioners gain a systematic understanding of the research developments in this field, thereby inspiring contributions to this exciting and important area. ### Key Features: - **Model-centric**: Discusses techniques that optimize model size for better performance. - **Data-centric**: Focuses on methods that improve efficiency by leveraging efficient data processing techniques. - **Framework-centric**: Covers advancements in frameworks that enhance the efficiency of LLMs. ### Tech Stack: - **Model-centric**: Techniques such as quantization, pruning, and knowledge distillation. - **Data-centric**: Methods like data parallelism, model parallelism, and efficient data storage. - **Framework-centric**: Frameworks that optimize resource usage, such as PyTorch, TensorFlow, and Hugging Face Transformers. ### Why It Matters: Efficient LLMs are crucial as they address the significant resource demands of large models, enabling their wider adoption and practical applications in various domains. This survey and repository aim to provide a valuable resource for researchers and practitioners to stay updated with the latest advancements in this field.

### Overview: This repository contains tutorials and projects on LangChain and Prompt Engineering for large language models, specifically focusing on the fine-tuning of Llama 2 and other models on custom datasets. The tutorials cover various aspects of using LangChain, including loading and indexing data, creating prompt templates, and using retrieval QA chains to query custom data. ### Key features: - Fine-tuning Llama 2 on a custom dataset - Loading custom data with loaders - Adding AI with models - Making LLMs useful with chains - Building chatbots with memory - Complex tasks with agents ### Tech stack: Python with Jupyter Notebook ### Why it matters: This repository provides practical, step-by-step tutorials and projects for developers looking to build real-world AI applications using ChatGPT/GPT-4 and LangChain. The tutorials cover essential topics such as fine-tuning models, loading custom data, and building chatbots, making it a valuable resource for anyone interested in leveraging large language models for practical applications.

### Overview: Windows Copilot API is a free, OpenAI-compatible API that converts the free chat experience of Microsoft Copilot into a RESTful API. It allows developers to access GPT-4 and GPT-5 models without the need for API keys or billing. ### Key features: - **Python Library:** Use the API as a Python library by calling `client.chat("Hi")`. - **Local OpenAI Server:** Run a local OpenAI-compatible API server at `http://localhost:8000/v1` that speaks the OpenAI format. - **Streaming and Conversations:** Supports token-by-token output and multi-turn conversations via `conversation_id`. ### Tech stack: Python ### Why it matters: It automates the consumer Copilot web experience for personal use, making it accessible without the need for API keys or credits.

### Overview: OpenLRM is an open-source implementation of Large Reconstruction Models, focusing on the efficient and accurate reconstruction of 3D models from images. It leverages the power of Blender and Hugging Face's models to achieve high-quality results with minimal computational resources. ### Key Features: - **Model Training**: OpenLRM supports training on multiple datasets including Objaverse and MVImgNet. - **Inference**: Supports both small and large models, with features ranging from 512 to 1024 dimensions. - **Models**: Available in versions 1.1 and 1.0, with weights trained on Objaverse and MVImgNet. - **Training Data**: Supports training on Objaverse and MVImgNet datasets. - **License**: The weights and models are released under the Apache 2.0 license. ### Tech Stack: Python ### Why it Matters: OpenLRM offers a practical and efficient solution for large-scale 3D reconstruction, making it accessible and useful for various applications such as architectural visualization, gaming, and more. Its open-source nature ensures that it can be easily integrated into existing projects and further developed by the community. ### Repository: 3DTopia/OpenLRM ### GitHub Stars: 1254

### Overview: Nautilus Compass is an open-source memory and reliability layer for AI agents, designed to enhance the performance and reliability of AI models. It beats mem0 on all three LongMemEval-S metrics while being fully local and 14× cheaper. ### Key features: - **Long-term memory that beats mem0 on all three LongMemEval-S metrics** - **Drift detection and cross-agent contracts that no other memory layer ships** - **Plugin for Claude Code / Desktop · Cline · Cursor · Continue.dev · Zed · any MCP client** ### Tech stack: Python ### Why it matters: - **Enhances AI model performance and reliability** - **Simplifies agent integration and ecosystem development** - **Offers a robust memory layer for AI agents** ### Repository: chunxiaoxx/nautilus-compass ### GitHub stars: 1254 ### GitHub description: Open memory & reliability layer for AI agents + independent judging/verification (NACRE judge, caliber-bench meta-benchmark, preregistered criteria) — by Nautilus Platform ### Primary language: Python ### GitHub description in English: This project powers the independent judging layer of the Nautilus platform, providing preregistered criteria, three-state verdicts, and a meta-benchmark for AI agents. It also offers a memory layer that beats mem0 on all three LongMemEval-S metrics while being fully local and 14× cheaper. It is available as a plugin for Claude Code, Cline, Cursor, Continue.dev, Zed,

### Overview: AttackGen is a cybersecurity incident response testing tool that leverages large language models and the MITRE ATT&CK framework to generate tailored incident response scenarios based on user-selected threat actor groups and organization details. ### Key Features: - **Custom Scenario Generation**: Tailored incident response scenarios based on chosen threat actor groups or ATLAS case studies. - **Organizational Tailoring**: Specifies organization size and industry for a tailored scenario. - **Framework Support**: MITRE ATT&CK Enterprise, ICS (Industrial Control Systems), and ATLAS (Adversarial Threat Landscape for AI Systems) frameworks. - **Detailed Techniques Display**: Lists the techniques used by the selected threat actor group or case study. - **Scenario Templates**: Supports common cyber incidents, including AI/ML-specific attack patterns. - **AI Insider Threat Scenarios**: Generates scenarios shaped by the agent's deployment archetype, threat model from an insider threat model. - **Chat Interface**: AttackGen Assistant for scenario updates and questions. - **Setup Guidance**: Guided setup with a consistent sidebar, summary of requirements, and Generate button disabled until workflow is ready. - **Persistent Results**: Results persist across reruns and navigation, with compact summaries and section navigation. - **User Feedback**: Captures feedback on scenario quality. - **Scenario Downloads**: Downloadable scenarios in Markdown format. - **OpenAI API Integration**: Supports multiple OpenAI-compatible endpoints for scenario generation. - **Docker Container**: Available as a Docker image for easy deployment. - **LangSmith Integration**: Optional integration with LangSmith for debugging, testing, and monitoring model

Overview: Self-Play Fine-Tuning (SPIN) is an innovative method that allows language models to improve themselves through self-generated training data, without requiring additional human-annotated preference data. This approach significantly enhances the performance of language models across various benchmarks. Key features: - SPIN utilizes a self-play mechanism, enabling LLMs to refine their policies by generating self-generated training data. - SPIN can outperform models trained with direct preference data, as demonstrated by its superior performance on the HuggingFace Open LLM leaderboard. - SPIN is designed to be easily reproducible and can be used to fine-tune models on various benchmarks. Tech stack: Python Why it matters: SPIN offers a unique and effective way to enhance language model performance, making it a valuable addition to the field of natural language processing.

### Overview: This repository contains a collection of must-read papers on knowledge editing for large language models, focusing on the advancements and challenges in knowledge management within these models. ### Key features: - **Knowledge Mechanism and Fusion for LLMs**: Discusses the mechanisms and techniques for integrating knowledge into large language models. - **Editing and Fusion for LLMs**: Explores the methods for modifying and integrating knowledge to enhance the model's performance. - **Knowledge Editing for LLMs**: Reviews the current state of knowledge editing in large language models and its impact on model performance. ### Tech stack: - Python, Jupyter Notebook, and other relevant tools for data analysis and research. ### Why it matters: - **Advancing Model Understanding**: Knowledge editing helps in improving the understanding and control of large language models. - **Enhancing Model Performance**: Techniques discussed in these papers can significantly enhance the performance of large language models by integrating and editing knowledge effectively. - **Research and Development**: These papers contribute to the ongoing research and development in the field of knowledge editing for large language models, providing valuable insights and benchmarks for future work. ### Repository: - **zjunlp/KnowledgeEditingPapers** - **GitHub Stars**: 1254 - **GitHub Description**: Must-read Papers on Knowledge Editing for Large Language Models.

### Overview: Paint by Example: Exemplar-based Image Editing with Diffusion Models is a project that leverages diffusion models to achieve precise and controllable image editing. The method uses self-supervised training to disentangle and reorganize source images and exemplar images, avoiding fusing artifacts. It introduces an information bottleneck and strong augmentations to ensure the controllability of the editing process and high fidelity results on in-the-wild images. ### Key features: - Exemplar-based image editing - Self-supervised training for disentanglement - Information bottleneck and strong augmentations - Controllable editing on high-fidelity images - Supports in-the-wild images ### Tech stack: Python ### Why it matters: Paint by Example demonstrates the potential of exemplar-based image editing with diffusion models, showcasing improved precision and controllability over traditional methods. The project's approach to avoiding fusing artifacts and ensuring high fidelity on real-world images makes it a significant contribution to the field of image editing.

### Overview: UForm is a pocket-sized multimodal AI library designed for efficient content understanding and generation across multilingual texts, images, and video clips. It offers compact, 64-dimensional embeddings for fast search and throughput, and comes with ONNX support for easy deployment on any platform. ### Key Features: - **Tiny Embeddings**: 64-dimensional Matryoshka-style embeddings for extremely fast search. - **Throughput**: 2-4x faster than competitors. - **Portable**: Models come with native ONNX support, making deployment easy. - **Quantization Aware**: Down-cast embeddings from `f32` to `i8` without losing much recall. - **Multilingual**: Trained on a balanced dataset, ensuring recall across over 20 languages. ### Models: - **uform3-image-text-english-large**: 365M parameters, 1 language, 12-layer BERT, ViT-L/14 - **uform3-image-text-english-base**: 143M parameters, 1 language, 4-layer BERT, ViT-B/16 - **uform3-image-text-english-small**: 143M parameters, 1 language, 4-layer BERT, ViT-B/16 ### Why it Matters: UForm is designed to be as efficient as it is versatile, making it suitable for a wide range of applications, from fast image captioning to advanced VQA. Its compact models and fast inference speed make it an ideal choice