Projects

Overview: This repository is a collection of resources and papers on Diffusion Models, providing a comprehensive overview of the field. Key features: - Introductory Posts - Introductory Papers - Introductory Videos - Introductory Lectures - Tutorial and Jupyter Notebook Tech stack: HTML Why it matters: This repository serves as a valuable resource for anyone interested in understanding and utilizing Diffusion Models in various applications, including generative modeling, image translation, and more. Repository: diff-usion/Awesome-Diffusion-Models GitHub stars: 12380

Overview: This repository contains a comprehensive reference guide for using MidJourney AI with various styles and keywords. It includes detailed information on image resolutions, weights, and more. Key features: - Styles and Keywords - Resolution Comparison - Image Weights - Additional Resources Tech stack: Markdown, HTML, and images. Why it matters: This guide is essential for users looking to enhance their MidJourney AI experiences by leveraging specific styles and keywords effectively. Repository: willwulfken/MidJourney-Styles-and-Keywords-Reference GitHub stars: 12313

### Overview: MOSS 是一款支持中英双语对话与工具调用的开源语言模型。本仓库提供 MOSS-003 系列模型、对话数据、部署与微调代码。 ### Key features: - MOSS-003 系列模型：包括基座模型、微调模型、插件增强模型和偏好模型。 - 多轮对话数据：覆盖有用性、忠实性和无害性三个层面，包含约110万条对话数据。 - 插件增强：支持搜索引擎、文生图、计算器和解方程等四种插件。 - 4bit和8bit量化版本：提供低资源推理版本。 ### Tech stack: Python ### Why it matters: MOSS 是一个强大的工具，能够帮助用户进行中英双语对话和工具调用，同时具备良好的事实性和安全性，以及稳定的回复质量。

### Overview: 本项目旨在为大模型应用开发者提供全面的RAG（检索增强生成）技术学习和实践指南。通过系统化学习和动手项目实践，帮助开发者掌握RAG技术的理论基础和实际应用技能，构建生产级的智能问答和知识检索系统。 ### Key features: - **RAG技术基础**：深入浅出地介绍RAG的核心概念、技术原理和应用场景。 - **数据处理全流程**：从数据加载、清洗到文本分块的完整数据准备流程。 - **索引构建与优化**：向量嵌入、多模态嵌入、向量数据库构建及索引优化技术。 - **检索技术进阶**：混合检索、查询构建、Text2SQL等高级检索技术。 - **生成集成与评估**：格式化生成、系统评估与优化方法。 - **项目实战**：从基础到进阶的完整RAG应用开发实践。 ### Tech stack: Python ### Why it matters: 随着大语言模型的快速发展，RAG技术已成为构建智能问答系统、知识检索应用的核心技术。本项目通过系统化学习和实践项目，帮助开发者掌握RAG技术的理论基础和实际应用技能，构建生产级的智能问答和知识检索系统。

### Overview: ARTEMIS is a powerful tool that turns natural-language instructions into reliable Android automation. It automates end-to-end workflows, captures logs, and integrates seamlessly with AI coding assistants such as Antigravity, Codex, and Claude Code. ARTEMIS achieves a 99%+ success rate on Google Research's AndroidWorld Benchmark. ### Key Features: - **Cross-App Automation**: Executes testing workflows and everyday tasks on Android from natural language instructions. - **Multimodal Targeting**: Uses element indices when available, with coordinate and visual locating fallbacks for custom interfaces. - **IDE Diagnostics**: Model Context Protocol (MCP) integration allows Antigravity, Claude Code, and Windsurf to drive test devices and collect Logcat output and screenshots. - **Flash Execution**: A reactive observe-and-act loop with asynchronous history summaries, typically 3–5 seconds per step. - **Pro Exploration**: Checks targets before individual actions and returns blocked actions to the Operator for recovery. Supports long-running exploratory and stability tests. - **AndroidWorld Results**: Achieves 99%+ task completion on Google Research's AndroidWorld benchmark (100+ multi-step tasks). ### Tech Stack: - Python 3.12+ ### Why It Matters: ARTEMIS simplifies the process of automating Android applications, making it easier for developers to test and maintain their apps. Its ability to integrate with AI coding assistants and achieve high success rates on benchmark tests makes it a valuable tool for anyone working on Android development projects.

Overview: This project is the official GitHub repository for the survey paper "A Survey of Large Language Models," authored by Zhao, Xin, Zhou, Kun, Li, Junyi, Tang, Tianyi, Wang, Xiaolei, Hou, Yupeng, Min, Yingqian, Zhang, Beichen, Zhang, Junjie, Dong, Zican, Du, Yifan, Yang, Chen, Chen, Yushuo, Chen, Zhipeng, Jiang, Jinhao, Ren, Ruiyang, Li, Yifan, Tang, Xinyu, Liu, Zikang, Liu, Peiyu, Nie, Jian-Yun, and Wen, Ji-Rong. The survey paper is available at [arxiv.org/abs/2303.18223](https://arxiv.org/abs/2303.18223). Key features: - A collection of papers and resources related to Large Language Models. - The survey paper "A Survey of Large Language Models" by Zhao, Xin, et al. - The Chinese book of the survey, focusing on providing explanations for beginners in the field of LLMs. - New content on long CoT reasoning, including reasoning patterns, data construction approaches, training methods, and scaling efforts. Tech stack: Python Why it matters: This project serves as a comprehensive resource for understanding the landscape of Large Language Models, including the survey paper and its associated content, making it valuable for researchers, students, and developers interested in the field.

### Overview: This project is a comprehensive guide to training a large language model (LLM) from scratch using PyTorch. The repository includes all the necessary components to train a model with up to a billion parameters, from downloading the required data to generating text. The project starts with a basic transformer model and gradually evolves into a modern, state-of-the-art model incorporating advanced techniques like rotary positions, RMSNorm, SwiGLU, and more. ### Key features: - Training a model from raw text to an aligned reasoning model. - Implementing a modern version of the same model with advanced training techniques. - All algorithms written in plain PyTorch. - Pretraining, fine-tuning, and evaluation of the model. - Support for both GPU and CPU training. ### Tech stack: - PyTorch ### Why it matters: This project is essential for developers and researchers interested in understanding the inner workings of large language models and experimenting with cutting-edge techniques. It provides a clear, step-by-step guide for anyone looking to train their own LLM from scratch, making it a valuable resource for both academic and practical applications. ### Repository: - **Repository**: FareedKhan-dev/train-llm-from-scratch - **GitHub stars**: 12168 - **Primary language**: Python - **GitHub description**: A straightforward method for training your LLM, from downloading data to generating text. - **GitHub description**: A pretraining tutorial that now goes all the way from raw text to an aligned, reasoning style model, with every algorithm hand-written in

Overview: Compromise is a natural language processing library designed to make text easy to parse and use, focusing on accuracy and efficiency rather than complexity. Key features: - Easy text to make - Limited and sensible decisions - Small, quick, and often good-enough - Small, quick, and often good-enough Tech stack: JavaScript Why it matters: Compromise aims to simplify the process of turning text into useful data, making it accessible and efficient for developers without requiring overly complex or advanced techniques.

Overview: Nerfstudio is a collaborative-friendly studio designed for creating, training, and testing NeRFs, providing a simplified end-to-end process with a more interpretable implementation of NeRFs through modular components. Key features: - Simplified API for easy integration - Modular NeRF implementation for user-friendly exploration - Community-friendly development environment - Comprehensive documentation and tutorials Tech stack: Python Why it matters: Nerfstudio aims to simplify the process of creating NeRFs, making it accessible and user-friendly for both beginners and experienced researchers, ultimately contributing to the broader field of NeRF technology.

### Overview: BISHENG is an open-source LLM DevOps platform designed for enterprise-level AI applications. It offers powerful features such as GenAI workflow, RAG, Agent, Unified Model Management, Evaluation, SFT, Dataset Management, Enterprise-level System Management, Observability, and more. ### Key Features: 1. **🚀 Task Mode:** - 🧠 Task Planning + Parallel Sub-Agents: Builds a task checklist, tracks progress, and assigns independent research tasks to sub-agents. - 🙋 Human in the Loop: Pauses the task when needed to gather information from the user. - 🧰 Plug-and-Play Skills: Supports PPT, Excel, and Word directly, or import skills from local files or GitHub. - 📎 Traceable Conclusions: Reports include citation markers, allowing users to open the source of the cited passages. 2. **📚 Knowledge Space:** - 🗂️ Organized Content: Multi-level folders, tags, card, and list views, with keyword and tag search capabilities. - 👥 Built by the Team: Invite team members with permission templates (Owner, Manager, Editor, Viewer) or publish the space to the Knowledge Square for others to join. - 🔐 Permissions: Access permissions can be granted to individuals, departments, or user groups in one step. Folders and files inherit permissions from their parent, and they can be set separately. - 💬 AI Q&A: Answers show sources, and clicking a source opens the original text at the cited

### Overview: MemOS is a Memory Operating System designed for LLMs and AI agents, providing unified memory management, multi-modal support, and enterprise-grade optimizations. It offers a single API for structured memory access, multi-cube KB management, asynchronous ingestion, and memory feedback correction. ### Key Features: - **Unified Memory API**: A single API for adding, retrieving, editing, and deleting memory. - **Multi-Modal Memory**: Supports text, images, tool traces, and personas. - **Multi-Cube Knowledge Base Management**: Manages multiple knowledge bases as composable memory cubes. - **Asynchronous Ingestion via MemScheduler**: Runs memory operations asynchronously for high production stability. - **Memory Feedback & Correction**: Refines memory with natural-language feedback. ### Tech Stack: TypeScript ### Why it Matters: MemOS offers a unified memory management system for LLMs and AI agents, improving context-aware interactions, multi-modal support, and enterprise-grade optimizations. It achieves 35.24% token savings and is leading in agent and user memory benchmarks.

### Overview: SpeechBrain is a PyTorch-based toolkit that accelerates Conversational AI development, focusing on speech and text processing tasks. ### Key features: - Comprehensive training recipes for over 200 tasks on 40 datasets. - Support for 20 speech and text processing tasks, including speech recognition, speaker recognition, and more. - Pretrained models on HuggingFace, facilitating seamless inference with just three lines of code. - Consistent code structure across different tasks for better replicability and reproducibility. ### Tech stack: Python ### Why it matters: SpeechBrain offers a holistic toolkit for advanced Conversational AI development, making it easier to create speech and text processing technologies.

### Overview: Allennlp is an open-source NLP research library built on PyTorch, designed for developing state-of-the-art deep learning models on a wide variety of linguistic tasks. ### Key features: - **PyTorch Backend**: Allennlp leverages PyTorch for efficient computation and model training. - **State-of-the-Art Models**: Supports a range of models and architectures for tasks like NER, text classification, and more. - **Modular Architecture**: Modules and neural network packages allow for easy integration and customization. - **Extensive Documentation**: Comprehensive documentation and guides for users and developers. ### Tech stack: Python ### Why it matters: Allennlp is a valuable tool for researchers and developers looking to build and experiment with advanced NLP models using PyTorch. Its modular architecture and extensive documentation make it a versatile choice for both academic research and practical applications.

### Overview: Ludwig is a declarative deep learning framework designed for building custom language models, neural networks, and tabular AI models. It simplifies the process of training, fine-tuning, and deploying AI models using a YAML config file and Python code. ### Key Features: - **Fine-tune LLMs with LoRA** - Fine-tune large language models like LLaVA, Qwen2-VL, and InternVL using LoRA adapters. - **Advanced PEFT adapters** - Support for initializers like PiSSA, EVA, CorDA/LoftQ, and adapter types like TinyLoRA, OFT, HRA, WaveFT, LN-Tuning, VBLoRA, and C3A. - **VLM fine-tuning** - Train models like LLaVA, Qwen2-VL, InternVL with `is_multimodal: true` and gated cross-attention. - **HyperNetwork combiner** - Conditionally generate weights for features to improve model performance. - **Nash-MTL & Pareto-MTL** - Multi-task loss balancing using game-theoretic and preference-based methods. - **LLM config generation** - Automatically generate YAML configs for LLMs using `ludwig generate_config`. - **ModelInspector** - Analyze model architecture, collect weights, and measure feature importance. - **Ray Serve & KServe** - Deploy models using Ray Serve and Kubernetes. - **GRPO alignment** - Reward-model-free RLHF via Group Relative Policy Optimization. - **torchao

Overview: This Python library provides a high-level API for semantic segmentation models, leveraging the power of PyTorch. It includes 12 encoder-decoder architectures, 800+ pretrained convolutional and transformer-based encoders, and popular metrics and losses for training routines. The library supports ONNX export and torch script/trace/compile friendly models. Key features: - Super simple high-level API - 12 encoder-decoder model architectures - 800+ pretrained convolution- and transform-based encoders, including support for `timm` - Popular metrics and losses for training routines - ONNX export and torch script/trace/compile friendly models Tech stack: Python Why it matters: This library simplifies the process of creating semantic segmentation models, making it accessible for developers and researchers alike. It provides a wide range of pretrained models and supports advanced features like ONNX export and script-friendly models, making it a valuable tool for image segmentation tasks. Repository: qubvel-org/segmentation_models.pytorch GitHub stars: 11,770 GitHub description: Semantic segmentation models with 500+ pretrained convolutional and transformer-based backbones.

### Overview: CocoIndex is an incremental engine designed for long-horizon agents, providing live, continuously fresh context for AI agents and LLM apps. It transforms codebases, meeting notes, inboxes, Slack, PDFs, and videos into a production-ready AI agent with always-updated context. ### Key Features: - **Incremental Processing**: Only the delta is reprocessed on every change, ensuring minimal processing overhead. - **Any Scale**: Works seamlessly across various environments and scales. - **Declarative**: Easy to use with a Python-based API, making it accessible for developers. - **10 Minutes Ready**: Quickly integrates into existing projects with minimal setup. ### Tech Stack: Rust ### Why It Matters: CocoIndex is essential for agents needing reliable, continuously fresh data, especially in enterprise environments where context is crucial for effective reasoning and decision-making.

### Overview: TensorZero is an open-source platform that unifies an LLM gateway, observability, evaluation, optimization, and experimentation. Designed for performance and ease of use, it integrates with various LLM providers and optimizes prompts and models for better performance. ### Key Features: - **Gateway:** Access every LLM provider through a unified API, with low latency (<1ms p99). - **Observability:** Store inferences and feedback in a database, accessible programmatically or in the UI. - **Evaluation:** Benchmark individual inferences or end-to-end workflows using heuristics, LLM judges, etc. - **Optimization:** Collect metrics and human feedback to optimize prompts, models, and inference strategies. - **Experimentation:** Ship with built-in A/B testing, routing, fallbacks, retries, etc. ### Tech Stack: Rust Rust is the primary language used for developing TensorZero. ### Why It Matters: TensorZero plays a crucial role in the AI ecosystem by providing a unified platform for integrating and optimizing various LLM providers, making it easier to experiment and optimize AI agents across diverse tasks.

Overview: LanceDB is an open-source, developer-friendly embedded retrieval library designed for efficient vector search and multimodal data management. It supports fast vector similarity search, full-text search, and SQL queries, making it ideal for AI/ML applications. Key features: - Fast vector search with state-of-the-art indexing - Comprehensive search capabilities including vector similarity, full-text, and SQL support - Multimodal support for storing, querying, and filtering vectors, metadata, and multimodal data - Advanced features like zero-copy, automatic versioning, and GPU support for building vector indexes Tech stack: Rust Why it matters: LanceDB provides a fast, scalable, and production-ready solution for vector search, making it an essential tool for developers building AI/ML applications. Its open-source nature and seamless integration with popular programming languages make it accessible and powerful for developers.

Overview: Rerun is a visual, queryable, and streamable data layer designed specifically for multimodal robotics data, enabling seamless ingestion, visualization, and training integration. Key features: - Multi-rate, multimodal data handling - Built-in viewer for real-time visualization - Data query and transformation capabilities - Integration with Python, Rust, and C++ SDKs - Streamable dataset mixes for training Tech stack: Rust on column-chunk storage Why it matters: Rerun simplifies the process of ingesting, querying, and visualizing physical AI data, making it easier to train models on complex multimodal datasets. Repository: [rerun-io/rerun](https://github.com/rerun-io/rerun) GitHub stars: 11,608

### Overview: HolaOS is an open-source agentic workspace designed to fit any company's workflow. It seamlessly connects the systems and chat tools you already use, alongside any models or agents you choose. With local-first architecture, your data never leaves your machines, making it perfect for businesses looking to streamline their operations without relying on external cloud services. ### Key Features: - **Local-first Architecture**: Data stays on your machines, ensuring privacy and control. - **100+ Integrations**: Connects with various systems and chat tools. - **MCP (Model-Driven Chat) Integration**: Seamless integration with models and agents. - **BYOK (Bring Your Own Key):** Supports any model or agent, including Claude Code, Codex, and BYOK. - **Clicks to Setup**: Easy setup in just a few clicks, not months. - **Real, Interactive Apps**: Apps open as real, interactive surfaces next to your agent, not just chat transcripts. - **Side-by-Side Design**: Apps and agents share the same screen, allowing you to step in and take over. - **Customizable**: You can change the workspace to suit your specific needs and workflows. ### Tech Stack: TypeScript ### Why It Matters: HolaOS offers a flexible and customizable workspace that adapts to your company's unique needs. By connecting your existing systems and chat tools, along with any models or agents you choose, HolaOS ensures that your work runs smoothly and efficiently. With local-first architecture, you maintain control over your data and workflows, making it ideal for businesses looking