Projects

Overview: AIcademy is a community offering free AI education, providing a friendly environment for learning. Key features: - Free AI education - Friendly community - Installation guide for Pnpm - Development server accessible at http://localhost:3000 Tech stack: TypeScript Why it matters: AIcademy offers valuable resources for those interested in AI education, fostering a supportive learning community.

### Overview: This repository contains the toolbox code for a vision system that participated in the Amazon Picking Challenge 2016. The system achieved 3rd and 4th place in the stowing and picking tasks, respectively. It utilizes multi-view RGB-D data and data-driven, self-supervised learning to accurately estimate the 6D pose of objects in cluttered environments. ### Key Features: - **RGB-D Realsense Sensor Drivers**: Standalone and ROS packages for handling the Realsense sensor data. - **2D Object Segmentation**: A deep learning ROS package for training and testing 2D object segmentation models. - **6D Pose Estimation**: A ROS package for estimating the 6D pose of objects using pre-scanned 3D models. ### Tech Stack: C++ ### Why it Matters: This system demonstrates the effectiveness of combining multi-view RGB-D data with self-supervised learning to solve challenging problems in warehouse automation, particularly in the context of the Amazon Picking Challenge. The approach can be applied to improve the accuracy and robustness of object recognition and pose estimation in various industrial settings.

Overview: This repository is a curated list of resources dedicated to Natural Language Processing (NLP) in Polish, including models, tools, datasets, and papers. Key features: - Polish text datasets - Models and embeddings - Language processing tools and libraries - Papers, articles, and blog posts Tech stack: Python, NLP libraries, and datasets. Why it matters: This repository serves as a comprehensive resource for anyone interested in Polish NLP, providing essential tools and datasets to enhance language processing capabilities in the Polish language.

### Overview: λ-RLM is a framework for long-context reasoning that replaces free-form recursive code generation with a typed functional runtime grounded in λ-calculus. Instead of letting the model write arbitrary recursive control logic during execution, λ-RLM executes a compact library of pre-verified combinators and uses neural inference only on bounded leaf subproblems. ### Key Features: - **Typed Recursive Long-Context Reasoning**: λ-RLM replaces free-form recursive code generation with a typed functional runtime grounded in λ-calculus. - **Pre-Verified Combinators**: Executes a compact library of pre-verified combinators. - **Neural Inference on Bounded Subproblems**: Uses neural inference only on bounded leaf subproblems. ### Tech Stack: Python ### Why It Matters: - **Improved Accuracy**: Across weak, medium, and strong model families, λ-RLM improves accuracy over standard RLM while substantially reducing latency. - **Stronger Formal Structure**: λ-RLM provides formal guarantees including termination, closed-form cost bounds, controlled accuracy scaling with recursion depth, and an optimal partition rule under a simple cost model. - **Structured Composition**: λ-RLM turns recursive reasoning into a structured functional program with explicit control flow, making it more reliable and predictable. ### Highlights: - **29 / 36 Wins**: λ-RLM outperforms standard RLM across model-task comparisons. - **Up to +21.9 Accuracy Points**: Average accuracy points across model tiers. - **Up to 4.1× Lower Latency**: Significantly reduces

### Overview: Speculative Decoding is an exploration into recent techniques for accelerating the text-to-semantic decoder in the Searp-TTS framework. This repository aims to implement and optimize speculative decoding strategies to potentially improve the performance of the text-to-semantic decoder in Searp-TTS. ### Key Features: - **Cache Hidden Layers**: Caching the hidden layer during speculative decoding to share resources between small and large models. - **Early Exit Strategy**: Implementing an early exit mechanism to reduce computational overhead. - **Batched Speculative Decoding**: Optimizing batched speculative decoding to improve performance. - **Prophet Transformer**: Optimizing speculative sampling with the Prophet Transformer idea. - **Micro-Optimizations**: Continuous micro-optimizations to further improve performance. ### Tech Stack: Python ### Why It Matters: Speculative Decoding is a technique that accelerates the decoding process of large language models by speculatively decoding ahead of the main decoding process. This approach can significantly reduce the computational cost of decoding, making it particularly useful in applications where performance is critical, such as in text-to-semantic models like Searp-TTS. The repository aims to provide a practical implementation of speculative decoding strategies, showcasing their effectiveness and efficiency.

### Overview: ImageFolder is a high-performance image tokenizer designed for Variational Autoencoders (VAR) and Autoregressive Models (AR). It offers robust tokenization and efficient image generation, making it suitable for various applications in AI and machine learning. ### Key Features: - **Robust Latent Matters**: Boosts image generation by reducing sampling error synthesis. - **XQ-GAN**: An open-source framework for image tokenization compatible with XQ-GAN. - **ImageFolder**: Supports autoregressive image generation with folded tokens, compatible with XQ-GAN. ### Tech Stack: - Python ### Why it Matters: ImageFolder is a powerful tool for generating high-quality images using autoregressive models, making it essential for applications in AI, image generation, and natural language processing.

Overview: This repository aggregates a vast collection of top-tier AI and language models papers, providing concise summaries in just 5 minutes. It covers a wide range of fields and conferences, including ACL, CVPR, NeurIPS, and more. Key features: - 23,000+ AI, LLM, NLP, CV, and related papers - Continuous updates across various conferences and research areas - Links to full papers and detailed lists Tech stack: Python Why it matters: This tool helps developers and researchers stay updated with the latest AI advancements in a quick and efficient manner.

### Overview: Gen-L-Video is an innovative method for generating and editing videos that can handle videos of any length and multiple segments with diverse semantic information. This approach extends short video diffusion models to generate and edit videos of hundreds of frames with multiple semantic segments, without requiring additional training. It addresses the limitations of current methodologies by providing a universal methodology that can be applied to any video length and format. ### Key Features: - **Universal Methodology**: Extends short video diffusion models to generate and edit videos of any length and format. - **Multi-Text Conditioning**: Generates videos with diverse semantic segments using a single text condition. - **Efficient Generation**: Preserves content consistency while generating videos of hundreds of frames. - **Versatile Applications**: Suitable for various video editing and generation tasks, including but not limited to, video editing, animation, and content creation. ### Tech Stack: - **Python**: The primary programming language used for implementing the project. - **Jupyter Notebook**: Used for interactive coding and documentation. ### Why it Matters: This methodology offers a significant advancement in video generation and editing by enabling the creation and modification of videos of any length and format, without the need for additional training. It addresses the limitations of current methodologies by providing a universal approach that can be applied to any video length and format, making it highly versatile and applicable to a wide range of video editing and creation tasks.

Overview: NanoDL is a JAX-based library designed specifically for training foundation models, particularly those under 4B parameters, for edge computing. It simplifies the process of creating and training transformer models from scratch, making it accessible even for AI/ML experts. Key features: - A wide range of transformer blocks and layers, allowing for the creation of custom models. - A variety of models, including Gemma, LlaMa3, Mistral, GPT3, GPT4 (inferred), T5, Whisper, ViT, Mixers, CLIP, and more. - Data-parallel distributed trainers for efficient training on multiple GPUs or TPUs. - Customizable dataloaders for streamlined data handling. - Layers not found in Flax/Jax, such as RoPE, GQA, MQA, and SWin attention. - GPU/TPU-accelerated classical ML models like PCA, KMeans, Regression, Gaussian Processes, and more. - True random number generators without verbose code. - Models contained in single files without external dependencies, making the source code easily reusable. - Advanced algorithms for NLP and computer vision tasks, such as Gaussian Blur, BLEU, Tokenizer, and more. Tech stack: Python with JAX, FLAX, and OPTAX. Why it matters: NanoDL addresses the challenge of developing and training transformer-based models efficiently, making it a valuable tool for AI/ML experts and researchers looking to create smaller-scale versions of these models for specific problems, while also providing a pedagogical approach to model development.

### Overview: LLM-Leaderboard is a joint community effort to create a central leaderboard for language models. Contributions and corrections are welcome. ### Key features: - Interactive Dashboard - Leaderboard with metrics for various LLMs ### Tech stack: Python ### Why it matters: LLM-Leaderboard serves as a centralized resource for benchmarking language models, facilitating comparisons and improvements in the field.

### Overview: CARLA is a Python library designed to benchmark and explore counterfactual explanation and recourse methods for machine learning models. It includes commonly used datasets and various machine learning models, making it a valuable resource for researchers and practitioners alike. ### Key Features: - **Common Datasets:** Adult, COMPAS, Give Me Some Credit, HELOC - **Machine Learning Models:** TensorFlow, PyTorch, Sklearn, XGBoost - **Counterfactual Methods:** Easily extendable with custom models and datasets - **Documentation:** Extensive documentation available [here](https://carla-counterfactual-and-recourse-library.readthedocs.io/en/latest/) ### Tech Stack: Python ### Why It Matters: CARLA provides a comprehensive toolset for understanding and evaluating various counterfactual explanation and recourse methods, ensuring researchers and practitioners can effectively analyze and improve their machine learning models.

### Overview: Phonon is a speech recognition engine developed by Fermion Research. It supports multiple models for transcribing speech files and live microphone recordings. The engine runs on various platforms including Apple silicon, CPUs, Windows, and NVIDIA GPUs via CUDA. It transcribes 25 languages and provides an OpenAI-compatible endpoint. ### Key features: - **Transcription**: Transcribes speech files and live microphone recordings. - **Models**: Supports Phonon-2.5, Phonon-2.5 English, Phonon-2, Phonon-1, Phonon-1 Micro, and Phonon-1 Big. - **Compatibility**: Runs on macOS, Linux, and Windows, including Apple silicon, CPUs, and NVIDIA GPUs. - **Accuracy**: Models are available on the Fermion Research website, with accuracy details provided. ### Tech stack: Python ### Why it matters: Phonon is a powerful tool for speech recognition, making it essential for applications requiring accurate speech-to-text conversion across various platforms and languages.

### Overview: Linformer is an implementation of a language model designed for PyTorch, specifically optimized for linear complexity and fixed sequence length. It addresses the limitations of Linformer by ensuring it can handle auto-regressive tasks and supports variable sequence lengths. ### Key Features: - **Auto-Regressive Support:** Linformer can be used for auto-regressive tasks, making it suitable for applications like language modeling. - **Variable Sequence Length:** It supports sequences of variable lengths, allowing for more flexible input lengths. - **Efficient Attention Mechanism:** Linformer achieves linear complexity, making it highly efficient for large-scale models. - **Reversible Network:** The model is designed to be reversible, which can be useful for certain types of tasks. ### Tech Stack: Python ### Why It Matters: Linformer has been put into production by Facebook, demonstrating its effectiveness in real-world applications. Its linear complexity and ability to handle variable sequence lengths make it a valuable tool in the field of natural language processing and beyond. ### Repository: lucidrains/linformer ### GitHub Stars: 307 ### GitHub Description: Implementation of Linformer for Pytorch

### Overview: OptaPy is an AI constraint solver for Python designed to optimize planning and scheduling problems, including the Vehicle Routing Problem, Employee Rostering, Maintenance Scheduling, Task Assignment, School Timetabling, Cloud Optimization, Conference Scheduling, Job Shop Scheduling, Bin Packing, and many more. ### Key Features: - **AI Constraint Solver**: OptaPy integrates with the OptaPlanner engine, providing advanced AI capabilities for solving complex planning problems. - **Python Implementation**: Utilizes Python for faster execution compared to Java or Kotlin, making it suitable for Python developers. - **Domain Modeling**: The domain consists of Problem Facts, Planning Entities, and Planning Solutions, with clear separation of concerns. - **Problem Facts**: Define constants and immutable facts that do not change during the planning process. - **Planning Entities**: Represent entities with planning variables that can be changed during the optimization process. - **Planning Solutions**: Define the facts and entities of the problem, including constraints and objectives. ### Tech Stack: Python ### Why it Matters: OptaPy is an essential tool for Python developers looking to solve complex planning and scheduling problems efficiently. Its integration with OptaPlanner makes it a powerful addition to the Python constraint solving ecosystem, offering advanced AI capabilities without the performance overhead of Java or Kotlin.

Overview: This repository is a curated collection of tools, utilities, resources, and tips for developers, designed to enhance the developer journey and boost productivity. Key features: - JavaScript utilities - TypeScript utilities - Java utilities - Kotlin utilities - PHP utilities - Python utilities - Angular utilities - React utilities - Vue utilities - Microfrontend - Monorepo - Android utilities - Ionic utilities - NodeJS utilities - Java utilities - Rust utilities - Python utilities - AI utilities - Git utilities - GraphQL utilities - Database utilities - Kubernetes utilities - Container utilities - Blockchain utilities - Bitcoin utilities - Ethereum utilities - Holochain utilities - Infrastructure utilities - Kubernetes utilities - Container utilities Tech stack: JavaScript Why it matters: This repository is a valuable resource for developers looking to improve their coding skills and boost their productivity. It provides a curated list of tools and utilities that are essential for various programming languages, frameworks, patterns, and development stages.

### Overview Kaizen is an AI-powered suite designed to automate tedious development tasks, allowing developers to focus on innovation and value delivery. By integrating seamlessly into your workflows, Kaizen empowers teams to streamline their development process, reclaim 40% of their development time, and enhance their productivity. ### Key Features - **🤖 AI-Powered Code Reviews**: Instant, insightful code reviews - **🧪 Intelligent Test Generation**: Automated end-to-end and unit tests - **🎨 UI Testing Automation**: Streamlined component testing - **🔬 Proactive Code Scanning**: Early issue detection - **📝 Automated Documentation**: Effortless codebase documentation - **🕵️ Smart AI Logger**: AI-powered application monitoring ### Why Choose Kaizen? - **⏰ Reclaim 40% of Development Time**: Automate repetitive tasks and focus on what truly matters - building great software. - **💪 Empower Your Team**: Provide developers with AI-powered tools to enhance their productivity and job satisfaction. - **💡 Drive Innovation**: Free up mental bandwidth for creative problem-solving and feature development. - **🔗 Seamless Integration**: Easily fits into your existing development workflow, minimizing disruptions and maximizing efficiency. ### Tech Stack Python ### Why It Matters Kaizen is a game-changer for software development, providing a solution that can significantly improve team productivity and innovation. By automating repetitive tasks, developers can focus on what truly matters, leading to better software and faster time to market.

### Overview: Parakeet.cpp is an ultra-fast and portable implementation of NVIDIA's Parakeet speech recognition models, built entirely in C++. It leverages Axiom, a lightweight tensor library with automatic Metal GPU acceleration, for efficient inference on Apple Silicon GPUs. The project is designed to be lightweight and memory-efficient, with FP16 support for significant memory reduction and a 96x speedup over CPU for 10-second audio. ### Key Features: - **Multiple Decoders**: Supports CTC greedy, TDT greedy, CTC beam search, and TDT beam search. - **Word Timestamps**: Provides per-word start/end times and confidence scores for all decoders. - **Beam Search + LM**: Incorporates optional ARPA n-gram language model fusion for beam search. - **Phrase Boosting**: Uses context biasing via token-level trie for domain-specific vocabulary. - **Batch Transcription**: Transcribes multiple files in a single batched encoder forward pass. - **VAD Preprocessing**: Silero VAD strips silence before ASR, and timestamps auto-remap. - **GPU Acceleration**: Metal via axiom's MPSGraph compiler for a 96x speedup on Apple Silicon. - **FP16 Inference**: Supports half-precision weights and compute for ~2x memory reduction. - **Streaming**: Supports EOU and Nemotron models with chunked audio input. - **Speaker Diarization**: Combines with ASR for speaker-attributed words with Sortformer. - **C API**: Provides a Flat `

Overview: Plug-and-play long-term memory system for LLM agents, designed for diverse environments including dialogue agents, knowledge-intensive QA, and web automation. Key features: - **Plug-in**: Enhance your agent with 6 lines of code. - **Memory**: Organizes experience into compact, reusable knowledge units, allowing minimal context overhead. - **Memory Types**: Three types: Semantic (facts, concepts), Procedural (workflows, procedures), and Episodic (interaction sequences). - **LLM Enhancement**: Uses LLMs for intelligent knowledge extraction, retrieval, and reasoning. - **Memory Compression and Evolution**: Supports updating and evolving the memory graph. Tech stack: Python Why it matters: PlugMem is a task-agnostic system that can be easily integrated into existing agent pipelines, providing a general memory backbone for diverse environments such as dialogue agents, knowledge-intensive QA, and web automation. It enables agents to recall relevant knowledge with minimal context overhead, making it a valuable tool for improving decision-making and performance in various applications.

Overview: ReEvo is a cutting-edge framework that leverages large language models (LLMs) to enhance hyper-heuristics, offering a more efficient and scalable approach to solving complex optimization problems. Key features: - Employs Reflective Evolution (ReEvo) for scalable heuristic generation. - Supports Neural Combinatorial Optimization (NCO), Genetic Algorithm (GA), Ant Colony Optimization (ACO), Guided Local Search (GLS), and Constructive Heuristics. - Addresses problems like Traveling Salesman Problem (TSP), Capacitated Vehicle Routing Problem (CVRP), Orienteering Problem (OP), Multiple Knapsack Problems (MKP), Bin Packing Problem (BPP), and Decap Placement Problem (DPP). - Integrates with various LLMs (e.g., OpenAI API, ZhiPu API, Llama API) for seamless integration. Tech stack: Python Why it matters: ReEvo provides a novel and efficient way to solve optimization problems by leveraging the strengths of large language models, making it a valuable tool for researchers and practitioners in the field of computational optimization.

### Overview: 本项目自动抓取并索引科学空间的文章元数据，按研究主题进行规则分类，方便在 GitHub 上快速浏览并跳转到原文。 ### Key features: - 自动抓取和索引科学空间文章 - 按研究主题分类文章 - 提供规则分类的 Markdown 文件 - 便于在 GitHub 上快速浏览和跳转 ### Tech stack: HTML ### Why it matters: 通过提供忠实于原文的 Markdown 文件，帮助开发者更高效地查找和阅读科学空间的文章，从而更好地理解和应用相关知识。