Projects

Overview: Kompute is a general-purpose GPU compute framework built on Vulkan, designed to support a wide range of cross-vendor graphics cards (AMD, Qualcomm, NVIDIA, and friends). It offers blazing fast performance, mobile-enabled capabilities, asynchronous operations, and optimized data processing for advanced GPU use cases. Key features: - Flexible Python module for C++ SDK optimizations - Asynchronous and parallel processing support through GPU family queues - Mobile-enabled with Android NDK examples across several architectures - Bring-your-own-Vulkan design to integrate with existing Vulkan applications - Explicit memory ownership and management for GPU and host - Robust codebase with 90% unit test coverage - Advanced use-cases in machine learning, mobile development, and game development Tech stack: C++ and Vulkan Why it matters: Kompute is backed by the Linux Foundation as a hosted project, making it a reliable and scalable solution for GPU computing across multiple vendors. Its flexible and efficient design makes it ideal for a wide range of applications requiring GPU acceleration. Repository: KomputeProject/kompute GitHub stars: 2578

### Overview: This repository is a curated collection of papers, code implementations, benchmarks, and resources related to Large Language Models (LLMs) post-training methodologies. It serves as a valuable resource for researchers, developers, and enthusiasts interested in understanding and advancing the state-of-the-art in post-training techniques for LLMs. ### Key Features: - **Survey Paper**: "LLM Post-Training: A Deep Dive into Reasoning Large Language Models" (Accepted in IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)). - **Corresponding Authors**: Komal Kumar, Tajamul Ashraf, Omkar Thawakar, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, Philip H.S. Torr, Fahad Shahbaz Khan, and Salman Khan. - **GitHub Stars**: 2578 - **Repository Description**: A curated collection of papers, code implementations, benchmarks, and resources related to Large Language Models post-training methodologies. ### Tech Stack: Python ### Why It Matters: This repository provides a comprehensive overview of the latest advancements in post-training methodologies for LLMs, including fine-tuning, reinforcement learning, and test-time scaling methods. It serves as a valuable resource for researchers, developers, and enthusiasts looking to stay updated with the latest developments in the field.

### Overview: Apache Burr (incubating) is a framework designed to simplify the development of applications that make decisions, such as chatbots, agents, and simulations. It integrates seamlessly with various frameworks and allows for easy integration with LLMs. The system includes a real-time telemetry UI for monitoring, tracking, and persisting application state, making it ideal for applications that require state management and complex decision-making. ### Key Features: - **Stateful Applications**: Express your application as a state machine, allowing for the management of complex state and decision-making. - **Easy Integration**: Works well with any framework and integrates with popular LLMs. - **Real-Time Monitoring**: A telemetry UI tracks and monitors the application in real-time, helping you understand and debug issues. - **Persistable State**: Supports pluggable persisters to save and load application state, ensuring data integrity and easy recovery. - **Chatbot Example**: A simple chatbot example is included to demonstrate how the UI can be used to interact with the application. ### Tech Stack: - Python ### Why It Matters: Apache Burr simplifies the development of applications that require state management, decision-making, and real-time monitoring. It provides a user-friendly interface for tracking and managing application state, making it an essential tool for developers working on complex applications that involve LLMs and stateful workflows. ### Repository: - Repository: `https://github.com/apache/burr` - GitHub Stars: 2575 - GitHub Description: "Build applications that make decisions (chatbots, agents, simulations, etc...

Overview: This repository contains advanced and agentic Retrieval-Augmented Generation (RAG) techniques, providing a comprehensive guide for researchers and developers looking to implement these methods effectively. Key features: - Implementation of advanced RAG techniques - Guidance on building and evaluating RAG systems - Clear explanations and code examples - Ready-to-use implementations for quick experimentation Tech stack: Jupyter Notebook Why it matters: This repository simplifies the process of implementing advanced RAG techniques, making it easier for researchers and developers to create effective RAG applications.

### Overview: 本项目是作者们根据个人面试和经验总结出的自然语言处理(NLP)面试准备的学习笔记与资料，该资料目前包含自然语言处理各领域的面试题积累。 ### Key features: - 命名实体识别 - 隐马尔科夫算法 - 最大熵马尔科夫模型 - 条件随机场（CRF） - 基于深度学习的命名实体识别方法 - 命名实体识别的词汇增强方法 - 命名实体识别的词汇/实体类型信息增强方法 - 命名实体识别的嵌套命名实体识别方法 - 命名实体识别的标注数据不足处理方法 - 命名实体识别的标注数据噪声问题 - 基于深度学习的命名实体识别方法的结构和优点 - 基于深度学习的命名实体识别方法的缺点 - 基于深度学习的命名实体识别方法的标签解码器介绍 ### Tech stack: Python, TensorFlow, Keras, Numpy, Scikit-learn, NLTK, SpaCy, PyTorch, etc. ### Why it matters: 该资料是NLP面试准备的重要参考资料，涵盖了面试中常见的命名实体识别、隐马尔科夫算法、最大熵马尔科夫模型、条件随机场等技术的面试题，对于准备NLP面试非常有帮助。

Overview: Big-Sleep is a simple command-line tool that allows users to generate images using natural language descriptions. It leverages OpenAI's CLIP and a BigGAN generator to create surreal and imaginative images. Key features: - Generate images using natural language input - Save images to the current directory - Support for multiple phrases in a single command - Option to penalize certain prompts Tech stack: Python Why it matters: Big-Sleep simplifies the process of generating images with complex prompts, making it accessible to a wide audience. It's particularly useful for artists, designers, and anyone looking to explore the boundaries of language and art. Repository: lucidrains/big-sleep GitHub stars: 2575 GitHub description: A simple command line tool for text to image generation, using OpenAI's CLIP and a BigGAN.

### Overview: This repository is a curated list of autonomous improvement loops, research agents, and autoresearch-style systems inspired by Karpathy's autoresearch. It includes a variety of platforms, hardware forks, domain-specific adaptations, evaluation methods, notable use cases, and related resources. The repository aims to provide a high-signal index for developers interested in improving their research and development processes. ### Key features: - General-purpose descendants: Frameworks and skills that can be applied to various domains. - Research-agent systems: Systems designed to automate research and experimentation. - Platform ports and hardware forks: Tools that can be adapted to different environments. - Domain-specific adaptations: Systems tailored to specific industries or domains. - Evaluation & benchmarks: Metrics and benchmarks for measuring the effectiveness of research. - Notable use cases and writeups: Examples of successful applications of these systems. - Related resources: Additional materials and links for further exploration. ### Tech stack: The repository uses a combination of Markdown, GitHub, and various open-source tools and frameworks to organize and present the information. ### Why it matters: This repository serves as a valuable resource for developers looking to enhance their research and development practices. By providing a curated list of tools and systems inspired by Karpathy's autoresearch, it offers a starting point for those interested in improving their research and development processes.

Overview: This project implements a Deep Neural Network (DNN) for real-time lane detection using the Lanenet model, based on the IEEE IV conference paper "Towards End-to-End Lane Detection: an Instance Segmentation Approach". The model is designed to be compatible with TensorFlow 1.12.0 and above, and has been tested on Ubuntu 16.04 with CUDA-9.0 and cudnn-7.0, requiring a GTX-1070 GPU. Key features: - Real-time lane detection using a DNN model - Implementation based on the IEEE IV conference paper - Compatible with TensorFlow 1.12.0 and above - GPU requirement: GTX-1070 - Model architecture details provided Tech stack: Python Why it matters: This project provides a practical solution for real-time lane detection, which is crucial for applications such as autonomous driving, traffic monitoring, and more. The implementation details and compatibility with TensorFlow make it a valuable resource for developers working on similar projects.

### Overview: This project is a collection of JavaScript libraries designed to interact with the Hugging Face API, providing TypeScript types for ease of use. The libraries include features such as interacting with the Hugging Face Hub, using various inference providers, and parsing and rendering OpenUSD files. ### Key Features: - **@huggingface/hub**: Interact with the Hugging Face Hub to create or delete repositories and commit/download files. - **@huggingface/inference**: Use all supported (serverless) inference providers or switch to inference endpoints to make calls to 100,000+ machine learning models. - **@huggingface/mcp-client**: A Model Context Protocol (MCP) client and a tiny agent library built on top of inference client. - **@huggingface/gguf**: A GGUF parser that works on remotely hosted files. - **@huggingface/dduf**: Similar package for DDUF (DDUF Diffusers Unified Format). - **@huggingface/lerobot**: A parser for LeRobot robotics datasets, reading v2.x and v3.0 layouts over HTTP requests. - **@huggingface/usd-web-viewer**: An OpenUSD viewer for the web, rendering `.usd`, `.usda`, `.usdc`, and `.usdz` files with three.js. - **@huggingface/tasks**: The definition files and source-of-truth for the Hub's main primitives like pipeline tasks, model libraries, etc. - **@huggingface/jinja**: A minimalistic JS

Overview: A developer-friendly skill for building a Karpathy-style knowledge base using Claude Code, Cursor, and Codex. It allows ingesting sources, compiling durable knowledge pages, answering questions with citations, and linting the wiki for consistency. Key features: - Ingesting sources into `raw/` - Compiling durable knowledge pages into `wiki/` - Answering questions with citations - Linting the wiki for index integrity, links, and health Tech stack: Python Why it matters: This skill provides a structured way to maintain and update knowledge bases using a LLM, making it easier to keep the wiki up-to-date and accurate. Repository: Astro-Han/karpathy-llm-wiki Primary language: Python GitHub stars: 2566 GitHub description: Agent Skills-compatible LLM wiki for Claude Code, Cursor, and Codex. Build a Karpathy-style knowledge base from raw sources, citations, and linting.

### Overview: 面向机器学习的特征工程是机器学习领域中至关重要的一环，它涉及如何有效地从原始数据中提取有用的特征，以便机器能够更好地理解和学习。本书详细介绍了特征工程的各个方面，从简单数字的奇特技巧到非线性特征提取和模型堆叠，再到自动化特征提取器的图像特征提取和深度学习，为读者提供了全面的指导。 ### Key features: - 简单数字的奇特技巧 - 文本数据：展开、过滤和分块 - 特征缩放的效果：从词袋到 TF-IDF - 类别特征：机器鸡时代的鸡蛋计数 - 降维：使用 PCA 压缩数据集 - 非线性特征提取和模型堆叠 - 自动化特征提取器：图像特征提取和深度学习 - 回到特征：将它们放到一起 ### Tech stack: JavaScript ### Why it matters: 本书为机器学习工程师提供了宝贵的指南，帮助他们从数据中提取出最有效的特征，从而提升模型的性能和准确性。通过本书，读者可以掌握各种特征工程技巧，无论是从文本数据到图像特征，都能找到适用的方法。

Overview: CubeStudio 是一款国产化、云原生的一站式开源人工智能平台，覆盖传统机器学习、深度学习与大模型全链路，提供多租户与算力纳管调度、算力租赁与 Token 中转站、多机多卡分布式训练、超参搜索、推理服务、vGPU 虚拟化、云边端协同与边缘计算、图文音多模态自动化标注、大模型 SFT 微调 / 奖励模型 / 强化学习训练、vLLM / Ollama / MindIE 大模型多机推理、私有知识库 / LLMOps / 智能体、AI 模型市场，以及从数据到上线的全流程模型部署。 Key features: - 支持多租户与算力纳管调度 - 多资源组/多集群支持 - GPU 虚拟化与禁用模型、共享模式、独占模式、vgpu 模式 - 支持多种算力，包括昇腾、寒武纪、海光、摩尔线程、沐曦等 - 支持算力市场租赁实例 - 支持多种网络配置，如公网/域名、反向代理和内网穿透方式访问、支持HTTPS - 支持国际化的多语言配置 - 支持数据标注，包括自动化标注与视觉大模型自动化标注 - 支持数据集管理，包括元数据库表管理、数据集一键探索

### Overview: Contrastive Unpaired Translation (CUT) is a novel method for image-to-image translation that leverages patchwise contrastive learning and adversarial training. This approach allows for unpaired image-to-image translation tasks without the need for hand-crafted loss functions or inverse networks, making it significantly faster and more memory-efficient compared to CycleGAN. The method can be extended to single image training, making it particularly useful for applications where only a single image is available. ### Key Features: - **Unpaired Image-to-Image Translation**: CUT provides a way to translate images from one domain to another without paired data, making it more practical for real-world applications. - **Patchwise Contrastive Learning**: Utilizes patchwise contrastive learning to enhance the translation quality by comparing patches between source and target images. - **Adversarial Training**: Incorporates adversarial training to further improve the translation quality by encouraging the model to generate more realistic outputs. - **Single Image Training**: Allows for single image training, reducing the need for paired images and making it suitable for applications where only a single image is available. ### Tech Stack: - **Python**: The implementation is written in Python, leveraging the PyTorch framework for deep learning. ### Why It Matters: - **Faster Training**: CUT is significantly faster and less memory-intensive compared to CycleGAN, making it more practical for real-world applications. - **Single Image Training**: The method can be extended to single image training, making it particularly useful for applications where paired images are not available. - **Versatility**: The ability to

Overview: tf2onnx is a tool that converts TensorFlow, Keras, Tensorflow.js, and Tflite models to ONNX format. Key features: - Converts TensorFlow, Keras, Tensorflow.js, and Tflite models to ONNX format - Supports ONNX opset versions 14 to 18 - Provides command-line and Python API options - Supports ONNX Runtime for runtime inference Tech stack: Python Why it matters: tf2onnx simplifies the conversion process for developers and researchers working with various TensorFlow-based models, making it easier to integrate these models into the ONNX ecosystem for further analysis, optimization, or deployment.

### Overview hCaptcha Challenger is a project designed to help developers face hCaptcha challenges with a multimodal large language model. It provides a seamless interface to face these challenges without relying on any tampermonkey scripts or third-party anti-captcha services. The project is built using Python and is open-source on GitHub, with a high number of stars indicating its popularity and usefulness. ### Key Features - **Image Labeling (Binary)**: Uses ResNet for binary classification. - **Image Labeling (Area Select: Point)**: Utilizes YOLOv8 for point detection. - **Image Labeling (Area Select: Bounding Box)**: Implements YOLOv8 for bounding box segmentation. - **Image Labeling (Multiple Choice)**: Applies ViT for zero-shot motion detection. - **Image Drag and Drop**: Uses Spatial Chain-of-Thought for this task. - **Ranking Strategy**: Supports nested-model-zoo for advanced ranking. - **Self-Supervised Challenge**: Implements CLIP-ViT for a self-supervised challenge. - **Agentic Workflow**: Employs AIOps Multimodal Large Language model for a more efficient workflow. ### Tech Stack The project is built using Python, leveraging the capabilities of a large language model to face hCaptcha challenges effectively. ### Why It Matters hCaptcha Challenger simplifies the process of facing hCaptcha challenges, making it easier for developers to integrate this functionality into their projects. By using a multimodal large language model, the project ensures that the challenges are handled efficiently and effectively,

Overview: EAGLE is an efficient method for fast decoding of Large Language Models (LLMs) that significantly boosts generation speed without compromising performance. This approach uses extrapolated contextual feature vectors to achieve remarkable improvements over vanilla decoding methods. Key features: - EAGLE is the fastest speculative method for LLMs, surpassing Lookahead and Medusa in speed. - EAGLE-2 improves upon EAGLE by using confidence scores for better performance. - EAGLE-3 removes the feature prediction constraint and simulates the process during training, enhancing generation speed and maintaining performance. Tech stack: Python Why it matters: EAGLE provides a practical solution for fast LLM decoding, making it accessible even to GPU-limited environments. The method's performance is backed by third-party evaluation, and it can be seamlessly integrated into various LLM serving frameworks, including ROCm, AngelSlim, AWS NeuronX Distributed Core, and CPM.c.

### Overview: OFA (OFA-Sys) is an official repository of the OFA (Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework) project. This project is designed to unify modalities and tasks through a simple sequence-to-sequence learning framework, supporting English and Chinese. OFA is used for tasks such as image captioning, visual grounding, visual question answering, text-to-image generation, text classification, text generation, image classification, and more. The repository provides step-by-step instructions for pretraining and finetuning, as well as corresponding checkpoints. ### Key Features: - **Image Captioning**: Supports English and Chinese. - **Visual Grounding**: Supports English and Chinese. - **Visual Question Answering**: Supports English and Chinese. - **Text-to-Image Generation**: Supports English and Chinese. - **Generic Interface**: Supports English and Chinese. - **Chinese OCR**: Supports Chinese text recognition. ### Tech Stack: - Python ### Why It Matters: OFA is a unified sequence-to-sequence pretrained model that can be used for a wide range of tasks and modalities. It simplifies the process of pretraining and finetuning by providing step-by-step instructions and corresponding checkpoints, making it easier for developers to integrate into their projects. ### Repository: - Repository: OFA-Sys/OFA - GitHub Stars: 2553 - GitHub Description: Official repository of OFA (ICML 2022). Paper: OFA: Unifying Architectures, Tasks, and Modal

Overview: DeepDetect is a deep learning runtime, command-line tool, and REST server designed for practical model operations, including creating services, training models, running predictions, monitoring jobs, and managing model repositories on the filesystem. Key features: - Supports images, text, CSV/tabular data, time series, sparse/SVM-style data, object-detection boxes, and segmentation masks through a single API surface. - Provides a Python wheel that embeds the DeepDetect runtime in the current Python environment and a CLI for repeatable model workflows. - Supports training and inference using the Torch and TensorRT frameworks. - Focuses on practical model operations, making it suitable for real-world applications. Tech stack: C++ and Python. Why it matters: DeepDetect simplifies the process of deploying and managing deep learning models, making it easier to create, train, and run models without the need for extensive knowledge of deep learning frameworks and environments. Repository: jolibrain/deepdetect GitHub stars: 2551

### Overview: Hivemind is a decentralized deep learning library built on PyTorch, designed to train large models across thousands of computers from various sources. This project aims to facilitate collaborative training of neural networks without a central authority, ensuring fault tolerance and efficient parameter aggregation. ### Key features: - Distributed training without a master node. - Fault-tolerant backpropagation. - Decentralized parameter averaging. - Training of neural networks of arbitrary size. ### Tech stack: Python ### Why it matters: Hivemind enables decentralized deep learning, allowing models to be trained on a global network of volunteers, making it ideal for distributed computing and enabling collaborative research without the constraints of a central authority.

Overview: AingDesk is a simple and easy-to-use AI assistant that supports knowledge bases, model APIs, sharing, web search, and intelligent agents, growing rapidly. Key features: - One-click deployment of local AI models and mainstream model APIs - Local knowledge base - Intelligent agent creation - Can be shared online for others to use - Supports web search - Supports server-side deployment Tech stack: TypeScript Why it matters: AingDesk is an easy-to-use AI assistant that supports local AI models, APIs, and knowledge base setup, growing rapidly.