Projects

Overview: fugashi is a Cython wrapper for MeCab, a Japanese tokenizer and morphological analysis tool. It provides fast, Pythonic Japanese tokenization and analysis, with wheels available for Linux, macOS, and Windows. Key features: - Fast Japanese tokenization and morphological analysis - Easy installation of UniDic dictionaries - Support for arbitrary dictionaries - Cross-platform compatibility Tech stack: C++ with Cython Why it matters: fugashi simplifies the process of tokenizing and analyzing Japanese text, making it easier for developers to integrate Japanese natural language processing into their projects. GitHub stars: 536 Repository: polm/fugashi

Overview: This project introduces a novel approach for implementing the KV cache in large language models, significantly reducing its memory footprint. Key features: - H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models - Optimizes the KV cache eviction policy for improved performance Tech stack: Python Why it matters: By reducing the memory footprint of the KV cache, this project enables more efficient deployment of large language models, particularly for applications requiring long-content generation.

Overview: This project provides a one-click portable installation of the AI-Toolkit by Ostris, a training suite for image and video diffusion models. The installer is designed to be easy to use without requiring system Python or VENV, making it ideal for developers and users on Windows systems with NVIDIA GPUs. Key features: - Installer: A single `.bat` file that starts the local server and opens the default browser once the server is live, and alerts for updates. - Launcher: A dedicated `.bat` file to start everything via one command, including an update feature. - Tech stack: Embedded portable version of Python, PyTorch, Triton, FlashAttention, torchcodec, and FFmpeg. - Why it matters: Simplifies the installation and setup process, making it accessible for developers and users on Windows with NVIDIA GPUs. Tech stack: Embedded portable version of Python, PyTorch, Triton, FlashAttention, torchcodec, and FFmpeg. GitHub stars: 532 Why it matters: Simplifies the installation and setup process, making it accessible for developers and users on Windows with NVIDIA GPUs.

Overview: SimilaritySearchKit is a Swift package that provides on-device text embeddings and semantic search functionality for iOS and macOS applications. It supports a variety of built-in and custom NLP models and metrics. Key features: - On-device text embeddings - Semantic search functionality - Built-in support for popular NLP models and metrics - Seamless integration for bring-your-own models - Swift package format Tech stack: Swift Why it matters: SimilaritySearchKit simplifies the development of applications that require text similarity and semantic search capabilities, ensuring privacy and performance without sacrificing functionality. Repository: ZachNagengast/similarity-search-kit Primary language: Swift GitHub stars: 536 GitHub description: 🔎 SimilaritySearchKit is a Swift package enabling *on-device* text embeddings and semantic search functionality for iOS and macOS applications. Emphasizing speed, extensibility, and privacy, it supports a variety of built-in state-of-the-art NLP models and similarity metrics, in addition to seamless integration for bring-your-own options. [](https://youtu.be/yYfQX4QdNJI) ## Use Cases Some potential use cases for **SimilaritySearchKit** include: - **Privacy-focused document search engines:**

### Overview: StoryTeller is a multimodal AI storyteller project that combines Stable Diffusion, GPT, and neural text-to-speech to create fully animated videos of short stories. It leverages the power of AI to generate a narrative and accompanying visuals, making it an excellent tool for storytelling enthusiasts and educators. ### Key Features: - **Story Generation**: Uses GPT to write the plot of a short story. - **Image Generation**: Draws an image for each sentence using Stable Diffusion. - **Voice Narration**: Uses neural text-to-speech (TTS) to narrate each line of the story. - **Customizability**: Supports customization of the beginning of the story with a prompt. ### Tech Stack: - **Stable Diffusion**: For image generation. - **GPT**: For generating the plot. - **Neural Text-to-Speech (TTS)**: For voice narration. ### Why It Matters: StoryTeller is a powerful tool for anyone looking to create engaging and visually stunning stories. It combines the creativity of AI with the beauty of art and the richness of language, making it a versatile and exciting project for both beginners and experienced storytellers.

### Overview: This repository contains the official PyTorch implementation of the paper "Towards Flexible Blind JPEG Artifacts Removal" presented at ICCV 2021. The model, FBCNN, is designed to remove JPEG artifacts from single and double JPEG images, as well as real-world JPEG images, achieving state-of-the-art performance on various quality factors. ### Key Features: - **Single JPEG Images (Color/Grayscale):** Removes artifacts from single JPEG images. - **Double JPEG Images (Aligned/Non-Aligned):** Handles double JPEG images with different alignment and quality factors. - **Real-World JPEG Images:** Applies to real-world JPEG images with varying quality and artifacts. ### Tech Stack: - **Python:** The primary programming language used. - **PyTorch:** The deep learning framework utilized for implementing the model. ### Why It Matters: This repository provides a practical solution for blind JPEG artifact removal, offering flexibility and convenience for real-world applications. It addresses the need for a model that can adapt to different quality factors, making it a valuable tool for image restoration and enhancement tasks.

### Overview: Parti-Pytorch is an implementation of Parti, Google's pure attention-based text-to-image neural network, in PyTorch. This project aims to provide a comprehensive and reproducible framework for training and using Parti models, leveraging the power of PyTorch for efficient and effective text-to-image generation. ### Key Features: - **Parti Model Implementation**: The project includes the complete implementation of Parti, a state-of-the-art text-to-image neural network. - **PyTorch Integration**: All components are written in PyTorch, ensuring compatibility and ease of integration with existing PyTorch projects. - **Training Code**: The repository also contains working training code for ViT VQGan VAE, a variant of the ViT architecture used in Parti, which can be used for further experimentation and replication. - **Modifications for Faster Training**: Additional modifications are included to improve training speed, based on literature from vision transformers. ### Tech Stack: The tech stack used in this project is primarily Python with PyTorch as the primary framework. It also includes T5 for text encoding and ViT for image generation, ensuring a comprehensive and efficient approach to text-to-image tasks. ### Why it Matters: Parti-Pytorch provides a practical and reproducible tool for researchers and developers interested in text-to-image generation. By using PyTorch, the project ensures that the implementation is both efficient and compatible with other PyTorch projects, making it a valuable resource for anyone looking to explore or extend the capabilities of Parti models.

Overview: This repository is a comprehensive knowledge base covering various aspects of computer science, including foundational topics like computer theory, computer technology, and security. It also delves into advanced topics such as AI, DevSecOps, and emerging technologies. Key features: - Computer Theory: Includes foundational topics like computer networks, operating systems, data structures, and algorithms. - Computer Technology: Covers language, frameworks, tools, and technical concepts. - Security: Includes security techniques, tools, research, and strategies for building and maintaining a secure environment. - Data: Discusses data structures, algorithms, and their applications. - AI: Covers AI algorithms, systems, and their applications. Tech stack: Python, Markdown, and GitHub. Why it matters: This repository serves as a valuable resource for developers looking to deepen their knowledge in computer science, security, and AI. It provides a comprehensive overview of foundational concepts and advanced topics, making it a go-to resource for both beginners and experienced developers. Repository: 404notf0und/Always-Learning GitHub stars: 535 GitHub description: 404 Not Found的知识库：计算机理论基础、计算机技术基础、底层研究、安全技术、安全研究、人工智能、企业安全建设、安全发展、职业规划、综合素质、国内外优秀技术人

### Overview: UniTok is a unified visual tokenizer designed for both generation and understanding tasks, compatible with various models such as LLaVA and Chameleon. It was accepted at NeurIPS 2025 as a spotlight paper and has been widely adopted for improving downstream tasks. ### Key features: - **Unified Tokenizer**: Designed for both generation and understanding tasks. - **Compatibility**: Works seamlessly with autoregressive generative models (e.g., LlamaGen) and multimodal understanding models (e.g., LLaVA). - **Unified MLLM**: Capable of multimodal generation and understanding using the Liquid framework, setting new state-of-the-art performance. ### Tech stack: - Python ### Why it matters: UniTok significantly enhances the performance of unified MLLMs in multimodal generation and understanding tasks, making it a game-changer in the field of visual language models.

### Overview: This repository is a comprehensive list of resources, features, scripts, and tutorials for better understanding and utilizing Kaldi, an open-source toolkit for speech recognition and speech processing. ### Key features: - **Kaldi Documentation**: Comprehensive documentation for beginners and advanced users. - **Speech Recognition Tutorials**: Detailed guides on how to start with speech recognition using Kaldi. - **Deep Learning Models**: Resources for training and developing deep neural network acoustic models with Kaldi. - **Speaker Diarization**: Tutorials and resources for speaker diarization using Kaldi. - **Graph Construction**: Scripts and tutorials for understanding and constructing decoding graphs in Kaldi. - **Custom Scripts**: Various scripts and utilities found in the Kaldi recipe directory, useful for debugging and model development. ### Tech stack: Kaldi, an open-source toolkit for speech recognition and speech processing. ### Why it matters: - **Comprehensive Resources**: Provides a wealth of information and tutorials to help users from beginners to advanced users. - **Community Contributions**: Encourages contributions by adding more links and resources, making it a collaborative project. - **Detailed Tutorials**: Offers detailed guides and tutorials that cover everything from basic concepts to advanced techniques. - **Scripts and Utilities**: Provides scripts and utilities that can help in various aspects of Kaldi usage, such as data manipulation and model development. ### Repository: YoavRamon/awesome-kaldi - **GitHub Stars**: 535 - **GitHub Description**: A list of features, scripts, blogs, and resources for better using Kaldi. -

### Overview: GraphRAG-rs is a high-performance, state-of-the-art Rust implementation of GraphRAG (Graph-based Retrieval Augmented Generation) that builds knowledge graphs from documents and enables natural language querying with configurable entity extraction and local LLM integration. ### Key features: - **Server-Only**: Supports Qdrant vector database and real embeddings via Ollama for production-ready applications. - **WASM-Only**: 100% client-side deployment for mobile and web applications. - **Hybrid**: A flexible hybrid model combining server and client-side components. - **GPU Acceleration**: Utilizes WebGPU for GPU acceleration, ensuring optimal performance. - **Local LLM Integration**: Configurable entity extraction using local LLMs for enhanced query results. - **CLI**: Simple command-line interface for quick start. - **Library**: Rust library for seamless integration into existing projects. ### Tech stack: Rust ### Why it matters: GraphRAG-rs offers a robust and efficient solution for building knowledge graphs from documents, enabling powerful natural language querying with advanced features like configurable entity extraction and local LLM integration. Its modular architecture and GPU acceleration support make it suitable for a wide range of applications, from SaaS to mobile and web-based solutions.

### Overview: Tabnine is an AI code assistant designed to enhance software development efficiency by providing real-time code completions, chat support, and natural language code completions across various programming languages and IDEs. ### Key features: - **Code Privacy:** Your code remains private and is never shared with Tabnine servers. - **Tabnine Chat (Beta):** A feature for asking questions and getting answers in real-time. - **Tabnine Pro:** Offers full-line full-function AI-powered code completions, including natural language to code completions. - **Tabnine Enterprise:** A self-hosted version for enterprise use, compatible with all JetBrains-based IDEs. ### Tech Stack: Kotlin ### Why it Matters: Tabnine significantly improves your development speed by reducing the time spent on repetitive tasks, thereby increasing productivity and overall efficiency in software development.

Overview: This project implements E(n)-Equivariant Graph Neural Networks (EGNNs) using PyTorch, focusing on invariant features and achieving state-of-the-art performance in dynamical system models and molecular activity prediction tasks. Key features: - Implementation of E(n)-Equivariant Graph Neural Networks using PyTorch - SOTA performance in dynamical system models and molecular activity prediction tasks - Supports both simple invariant features and complex equivariant features - Includes features for dynamic systems, molecular activity prediction, and equivariant feature extraction Tech stack: Python Why it matters: This project leverages PyTorch for efficient computation and offers state-of-the-art performance in complex graph neural network tasks, making it valuable for researchers and developers working on dynamical system models and molecular activity prediction. Repository: lucidrains/egnn-pytorch GitHub stars: 535 GitHub description: Implementation of E(n)-Equivariant Graph Neural Networks, in Pytorch. May be eventually used for Alphafold2 replication.

Overview: Sarathi-Serve is a low-latency and high-throughput LLM serving engine designed for efficient inference of large language models. Key features: - High throughput - Low latency - Optimized for LLM inference Tech stack: Python Why it matters: Sarathi-Serve is crucial for accelerating the performance of large language models, making it an essential tool for developers and researchers working in the field of artificial intelligence and natural language processing. Repository: microsoft/sarathi-serve Primary language: Python GitHub stars: 534 GitHub description: A low-latency & high-throughput LLM serving engine for LLMs README excerpt: Sarathi-Serve is a high throughput and low latency LLM serving framework. Please refer to our [OSDI'24 paper](https://www.usenix.org/conference/osdi24/presentation/agrawal) for more details.

### Overview: Mitsu is an AI-assisted penetration testing agent designed to enhance offensive security workflows by providing structured attack guidance, contextual analysis, and multi-backend integration. It streamlines the pentest process from reconnaissance to post-exploitation, using a chibi red-team companion to keep track of targets, IPs, and other relevant information. ### Key features: - **Model Providers:** Mitsu integrates with OpenAI, Anthropic, OpenRouter, Ollama, DeepSeek, Moonshot (Kimi), NVIDIA NIM, and Cloudflare Models, as well as custom endpoints compatible with OpenAI and Anthropic. - **Mitsu’s MCP Server:** Connects models to a catalog of 165+ Red Team Armory tools across 17 categories, enabling comprehensive engagement. - **Session Intel:** Tracks targets, IPs, domains, services, vulnerabilities, findings, accounts, evidence, and observations, and updates as the engagement progresses. - **Interactive Attack-Surface Map:** Converts session Intel into an interactive map for visual representation. - **Assessment Report:** Generates a detailed report summarizing findings and observations. - **Red Team Armory Integration:** Seamlessly integrates with Red Team Armory tools for a cohesive workflow. - **Custom Command Wrappers:** Allows for the creation of custom command wrappers for additional functionalities. ### Tech Stack: Python ### Why it Matters: Mitsu simplifies and streamlines the pentest process, making it easier and more efficient for security professionals to conduct comprehensive and effective penetration testing. Her AI-driven approach enhances the accuracy and effectiveness of the engagement

Overview: HardNet is a PyTorch implementation of the descriptor learning model described in the NIPS 2017 paper "Working hard to know your neighbor's margins: Local descriptor learning loss". This project includes the source code, examples, and pre-trained models for evaluating and using the HardNet descriptor. Key features: - HardNet implementation in PyTorch - Pre-trained models available in the repository - Examples for converting HardNet to Torchscript for use in C++ code - Description of improvements and updates to the model - Comparison with other descriptor learning models on various datasets Tech stack: Python Why it matters: HardNet is a valuable tool for computer vision tasks that require local descriptor learning, providing a robust and effective method for feature extraction and retrieval. Repository: DagnyT/hardnet GitHub stars: 534 GitHub description: Hardnet descriptor model implementation for NIPS 2017 paper.

Overview: This repository contains a curated collection of 3000+ high-quality prompts for the Gemini Nano Banana Pro, featuring 3200+ selected examples across various fields such as photography, design, art, and marketing. The prompts are sourced from Twitter, small bookmarked boards, and other platforms, ensuring authenticity and quality. The project includes a dual-language interface and automatic updates, making it an invaluable resource for creative professionals. Key features: - 3200+ Prompts - 900+ Creators - 3000+ Examples - Dual-language support - Automatic updates - JSON format Tech stack: JavaScript, TypeScript, React, and Node.js Why it matters: This project serves as a comprehensive resource for AI art enthusiasts, providing a vast array of prompts and examples that can be used to enhance creativity and innovation in various fields. The project's dual-language support and automatic updates ensure that users can access the latest and most relevant content, making it an essential tool for anyone looking to explore the potential of AI art.

### Overview: ModelDepot/tfjs-yolo-tiny is a JavaScript library that enables real-time object detection in web browsers using Tiny YOLO, a lightweight version of the YOLO object detection model. This implementation is optimized for performance and accuracy, making it suitable for applications requiring fast and accurate object detection in the browser. ### Key Features: - **Tiny YOLO Model Support**: Utilizes the Tiny YOLO model for efficient object detection, reducing computational complexity. - **Real-Time Object Detection**: Detects objects in images within milliseconds, making it ideal for use in web applications. - **In-Browser Implementation**: Runs directly in the browser, eliminating the need for complex server-side processing. - **Customizable Options**: Offers advanced configuration options such as filtering thresholds and class-specific settings. ### Tech Stack: - **JavaScript**: The primary language for development. - **Tensorflow.js**: Framework used to implement the Tiny YOLO model and perform object detection. ### Why It Matters: - **Performance**: Delivers high-speed object detection, making it suitable for applications requiring real-time performance. - **Ease of Use**: Provides a straightforward API for integrating object detection into web applications. - **Compatibility**: Works seamlessly with existing web technologies, ensuring compatibility with modern browsers. ### GitHub Stars: - **GitHub Stars**: 534 - **Repository**: ModelDepot/tfjs-yolo-tiny - **Primary Language**: JavaScript - **GitHub Description**: In-Browser Object Detection using Tiny YOLO on Tensorflow.js - **GitHub URL**:

### Overview: Reasoning on Graphs (RoG) is an official implementation of the ICLR 2024 paper "Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning." This project combines large language models (LLMs) with knowledge graphs (KGs) to enable faithful and interpretable reasoning. It provides a framework for generating relation paths, retrieving valid reasoning paths from KGs, and conducting faithful reasoning with LLMs. ### Key Features: - **Relation Path Generation**: RoG generates relation paths grounded by KGs as faithful plans. - **Knowledge Graph Retrieval**: These plans are used to retrieve valid reasoning paths from KGs for LLMs to conduct faithful reasoning. - **Interpretable Results**: The project ensures that the reasoning process is faithful and interpretable, making the results understandable and verifiable. ### Tech Stack: - Python ### Why it Matters: RoG leverages the strengths of large language models and knowledge graphs to enhance reasoning capabilities, making it a valuable tool for applications requiring faithful and interpretable knowledge-based reasoning.

### Overview: This repository provides an overview of all components used for the creation of BLOOMZ, mT0, and xP3 introduced in the paper "Crosslingual Generalization through Multitask Finetuning" by Samuel Albanie. The repository includes datasets, models, and evaluation metrics for cross-lingual generalization tasks. ### Key Features: - **Datasets:** - `xP3`: A mixture of 13 training tasks in 46 languages with English prompts. - `xP3mt`: A mixture of 13 training tasks in 46 languages with prompts in 20 languages (machine-translated from English). - `xP3all`: xP3 + additional 3 tasks for a total of 16 tasks in 46 languages with English prompts. - `xP3megds`: Megatron-DeepSpeed processed version of xP3. - `xP3`: Repreprocessed version of the English-only `xP3` with 8 training tasks. - `xP3small`: A smaller version of `xP3` with 300M parameters. - `xP3base`: A base version of `xP3` with 580M parameters. - `xP3large`: A large version of `xP3` with 1.2B parameters. - `xP3xl`: An even larger version of `xP3` with 3.7B parameters