Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 801 stars

vandit15/Class-balanced-loss-pytorch

Pytorch implementation of the paper "Class-Balanced Loss Based on Effective Number of Samples"

### Overview: This Pytorch implementation addresses the issue of class imbalance by proposing a loss function specifically designed for handling imbalanced datasets. The method calculates the effective number of samples for each class, which is then used to define a loss function that mitigates the impact of class imbalance. ### Key Features: - **Class-Balanced Loss Function**: Implements the Class-Balanced Loss method from the paper "Class-Balanced Loss Based on Effective Number of Samples" presented at CVPR'19. - **Effective Number of Samples Calculation**: Computes the effective number of samples for each class, which is crucial for balancing the loss across different classes. - **Pytorch Integration**: The implementation is fully compatible with Pytorch, making it easy to integrate into existing projects. ### Tech Stack: Python (>=3.6), Pytorch (>=1.2.0) ### Why it Matters: This method is particularly useful in scenarios where class imbalance significantly affects the performance of machine learning models, ensuring more accurate and reliable predictions.

Jupyter Notebook 800 stars

FareedKhan-dev/train-deepseek-r1

Building DeepSeek R1 from Scratch

### Overview: Building DeepSeek R1 from Scratch ### Key features: - **Code Implementation**: Detailed Jupyter notebook for step-by-step implementation. - **DeepSeek R1 Explanation**: Non-tech folks' guide to DeepSeek R1 architecture and training process. - **Base Model**: Utilizes a tiny base model to simplify the training process. - **Policy Model (R) in RL Setup**: Reinforcement Learning setup with a policy model (R). - **GRPO Algorithm for R1 Zero**: Gradient Policy Optimization for R1 Zero. - **Prompt Template**: Template for creating prompts. - **Preprocessing Training Data**: Techniques for preprocessing the training data. - **Reward Functions**: Various reward functions used in the training process. - **Training Configurations for R1 Zero**: Detailed configurations for training R1 Zero. - **Saving Tiny R1 Zero LLM**: Saving the tiny R1 Zero language model. - **Preparing Cold Start Data for SFT**: Preparing data for subsequent stages. - **Few-shot Prompting with Long CoT**: Using few-shot prompting with long context templates. - **Direct Prompting**: Direct prompt-based training. - **Post Processing Refinement**: Refinement steps after training. - **SFT Stage 1 Training**: Stage 1 of the SFT training loop. - **Stage 1 SFT Trainer Configs for R1**: Configurations for the SFT trainer for R1. - **Stage 1 STF Training Loop**: The training loop for Stage 1 of the SFT training. - **Saving Tiny R1

Python 800 stars

songweige/rich-text-to-image

Rich-Text-to-Image Generation

Overview: This project aims to enhance text-to-image generation by incorporating rich text formatting information, such as font size, color, style, and footnotes, to control and refine the text-to-image process. Key features: - Support for various rich text formats (e.g., HTML, Markdown) to input text prompts. - Integration with Stable Diffusion models (v1-5, SDXL, AnimeXL) for high-quality image generation. - Automatic conversion of rich text to JSON format for easy integration into applications. - Region-based diffusion for precise control over image synthesis. Tech stack: Python, PyTorch, Stable Diffusion models, Hugging Face. Why it matters: This approach provides greater control over text-to-image generation by leveraging rich text formatting, leading to more expressive and detailed images.

Python 800 stars

usemoss/moss

The retrieval layer for production AI systems. Lightning-fast (<10ms) search without vector databases. Built for browser, edge, on-device, and cloud.

Overview: Moss is a sub-10 ms semantic search runtime designed for Conversational AI agents. It leverages hybrid retrieval, built-in embeddings, metadata filtering, and a WebAssembly build, all integrated into a single SDK that can be embedded in applications. Key features: - Hybrid retrieval for real-time conversations - Built-in embeddings for semantic search - Metadata filtering for enhanced relevance - WebAssembly build for fast performance - SDK compatible with Python, TypeScript, Elixir, and C Tech stack: Python Why it matters: Moss is crucial for Conversational AI agents that need to respond quickly and accurately in real-time, ensuring a seamless user experience without the latency issues associated with remote vector databases.

Python 800 stars

weichen582/RetinexNet

A Tensorflow implementation of RetinexNet

Overview: RetinexNet is a Tensorflow implementation of RetinexNet, a method for low-light image enhancement. It was presented at the BMVC'18 conference and is based on the research of Chen Wei et al. Key features: - Deep Retinex Decomposition for Low-Light Enhancement - Implements Tensorflow - Paper: "Deep Retinex Decomposition for Low-Light Enhancement" by Chen Wei et al., BMVC'18 - Project Page & Dataset: [Daooshee's Project Page](https://daooshee.github.io/BMVC2018website/) - Requirements: Python, Tensorflow >= 1.5.0, numpy, PIL - Testing Usage: Run `python main.py --phase=test` to test your own images, or `python main.py` for demo cases - Training Usage: Download training data sets, set parameters, and run `python main.py` for training - Citation: Chen Wei, Wenjing Wang, Wenhan Yang, Jiaying Liu. "Deep Retinex Decomposition for Low-Light Enhancement." British Machine Vision Conference, 2018 Tech stack: Python Why it matters: RetinexNet is a significant contribution to low-light image enhancement, offering a practical solution for improving the quality of low-light images.

Python 799 stars

NVIDIA-AI-Blueprints/rag

This NVIDIA RAG blueprint serves as a reference solution for a foundational Retrieval Augmented Generation (RAG) pipeline.

Overview: The NVIDIA RAG Blueprint is a reference solution and foundational starting point for building Retrieval-Augmented Generation (RAG) pipelines using NVIDIA NIM microservices. It enables enterprises to deliver natural language question answering grounded in their own data, meeting governance, latency, and scalability requirements. Key features: - Agentic RAG: A LangGraph plan-and-execute pipeline for multi-hop, ambiguous, and cross-document queries, with scope discovery, parallel task execution, synthesis, and optional verification. - Data Ingestion: Multimodal content extraction from documents with text, tables, charts, infographics, and audio, and support for custom metadata. - Search and Retrieval: Agentic RAG pipeline with plan-and-execute retrieval, multi-collection searchability, hybrid search with dense and sparse search, reranking, GPU-accelerated Index creation and search, and pluggable vector database. - Query Processing: Query decomposition and dynamic filter expression creation. - Generation and Enrichment: Opt-in for Multimodal and Vision Language Model Support, document summarization with flexible page filtering and real-time progress tracking, and optional programmable guardrails for content safety. Tech stack: Python. Why it matters: The NVIDIA RAG Blueprint serves as a flexible starting point for developers to adapt and extend to their specific needs, providing an enterprise-ready framework for building Retrieval-Augmented Generation pipelines with NVIDIA NIM microservices.

Unknown 799 stars

affige/genmusic_demo_list

a list of demo websites for automatic music generation research

### Overview: This repository contains a collection of demo websites showcasing the latest advancements in automatic music generation research. The projects range from song generation and cover song creation to streaming and live generation, demonstrating the capabilities of various techniques and models in the field. ### Key Features: - Song Generation: Various models and techniques including YuE2, SongCraft, FullDiT, WanSong v1.0, LeVo 2, Khala, ComposerFlow, ACE-Step, Muse, DiffRhythm 2, JAM, SongBloom, LeVo, ACE-Step, VersBand, MusiCoT, DiffRhythm, SongGen, YuE, Melodist, SongCreator, MelodyLM, MuLaCover, MPEcho, SongEcho, SiMe, liveband, stream-music-gen, LatCH, AudioMoG, Stemphonic, TVC-MusicGen, and MuseControlLite. - Cover Song Generation: MuLaCover, MPEcho, SongEcho. - Streaming/Streaming Live Generation: SiMe, liveband, stream-music-gen. - Text-to-Music: LatCH, AudioMoG, Stemphonic, TVC-MusicGen. ### Tech Stack: Python, JavaScript, React, TensorFlow, PyTorch, Hugging Face, and various custom models and libraries. ### Why It Matters: These projects showcase the progress and innovation in automatic music generation, providing valuable insights into the current state of the field. They demonstrate the effectiveness of different techniques and models, highlighting the importance of research in this area

Markdown 798 stars

amitshekhariitbhu/ai-system-design

AI System Design - Learn how to design AI systems built on LLMs, RAG, and AI Agents step by step.

### Overview: This guide is designed to help anyone interested in AI system design, from beginners to advanced practitioners. It covers everything from foundational concepts to advanced techniques, including LLMs, GPUs, RAG, AI agents, and more. The goal is to provide a comprehensive understanding of AI system design, enabling you to confidently tackle AI-related challenges. ### Key Features: - **LLM Inference:** Learn how to design AI systems using Large Language Models (LLMs). - **GPU and TPU Usage:** Understand how to leverage GPUs and TPUs for inference and training. - **RAG Integration:** Explore the integration of Rich Argumentation and Generation (RAG) in AI systems. - **AI Agents and Multi-Agent Systems:** Discover how to design AI agents and multi-agent systems. - **Cost Optimization and Observability:** Learn best practices for cost optimization and observability in AI systems. - **Interview Preparation:** Get tips and strategies for cracking AI system design interviews. ### Tech Stack: - Markdown ### Why It Matters: This guide is essential for anyone looking to build and design AI systems. Whether you're a beginner or an experienced engineer, this guide will help you understand the intricacies of AI system design and how to implement them effectively. ### Repository: amitshekhariitbhu/ai-system-design ### GitHub Stars: 736 ### GitHub Description: AI System Design - Learn how to design AI systems built on LLMs, RAG, and AI Agents step by step. ### Table of Contents: - **About This

Go 798 stars

bruin-data/dac

DaC is a dashboard-as-code tool. Build interactive dashboards using YAML and JSX. Built-in semantic layer. Get your agents to build standardized, reviewable dashboards.

### Overview: Dac is a dashboard-as-code tool that allows developers to create interactive dashboards using YAML and JSX. It supports various databases and provides a built-in semantic layer for defining and referencing metrics and dimensions. The tool is designed to be reliable and reviewable, making it suitable for AI agents to build dashboards in a standardized and maintainable manner. ### Key Features: - **Dashboard-as-Code**: Define and validate dashboards using YAML and JSX. - **Built-in Semantic Layer**: Define metrics and dimensions in `semantic/` and reference them in any widget. - **Supports Major Databases**: Integrates with Bruin for querying databases like Postgres, MySQL, Snowflake, BigQuery, Redshift, Databricks, and more. - **Dynamic Charts, Tabs, Loops, and Conditionals**: Utilizes TSX for dynamic elements. - **Reliable and Reviewable**: Designed for AI agents to build dashboards in a secure and maintainable way. ### Tech Stack: Go ### Why It Matters: Dac simplifies the process of creating and maintaining dashboards, ensuring that developers can focus on their core responsibilities while leveraging the power of AI for data visualization. The tool's ability to integrate with various databases and its built-in semantic layer make it a valuable asset for teams looking to standardize and review their dashboards. ### Repository: bruin-data/dac ### GitHub Stars: 796 ### GitHub Description: DaC is a dashboard-as-code tool. Build interactive dashboards using YAML and JSX. Built-in semantic layer.

Python 797 stars

Optim-Agent/optim-agent

LLM agents as your hyperparameter optimizer.

Overview: optim-agent is an agent-based system optimization tool that leverages Claude Code or Codex to automatically tune real system parameters based on the code and observed outcomes. Key features: - Semantic proposals: coding agents reason over parameter meanings, context, and observed outcomes. - Small-budget leverage: useful when evaluations are expensive and classical surrogates are still data-starved. - Agent CLI upside: proposal quality can improve as the underlying coding agents improve. - Auditable decisions: JSON/SQLite studies retain configurations, outcomes, states, context, and optional agent rationale. - Bounded execution: the agent only proposes values; optim-agent validates them against the declared space, and invalid output falls back to safe sampling. Tech stack: Python Why it matters: optim-agent combines Claude Code or Codex with agent-based optimization to automate the parameter tuning process, making it more efficient and effective, especially when dealing with expensive evaluations or data-starved classical surrogates.

Python 797 stars

SlapBot/stephanie-va

Stephanie is an open-source platform built specifically for voice-controlled applications as well as to automate daily tasks imitating much of an virtual assistant's work.

Overview: Stephanie is an open-source platform designed specifically for voice-controlled applications and automating daily tasks, inspired by virtual assistant capabilities. Key features: - Voice-controlled applications - Automating daily tasks - Mimicking virtual assistant work Tech stack: Python Why it matters: Stephanie provides a comprehensive solution for automating tasks and enhancing daily productivity through voice commands, making it an invaluable tool for those looking to streamline their workflows and simplify their lives. Repository: SlapBot/stephanie-va Primary language: Python GitHub stars: 797 GitHub description: Stephanie is an open-source platform built specifically for voice-controlled applications as well as to automate daily tasks imitating much of a virtual assistant's work.

JavaScript 797 stars

markfulton/ai-employees

Open source AI Employees. 8 scheduled business roles, 60 routines, on Claude Code and 10 other harnesses. They drive your browser the way you do and improve every run. You own the files.

### Overview: AI Employees is an open-source project that provides a collection of 8 AI agents designed to run business roles, each with its own set of routines. These agents are compatible with various Claude Code and other harnesses, making them versatile and easy to integrate into any environment. ### Key Features: - **8 AI Agents**: Each agent runs a specific business role, such as a Sales Representative, Marketing Manager, or Customer Support Specialist. - **60 Routines**: Each agent has a dedicated set of routines that automate tasks related to its role. - **Compatible with Claude Code and Other Harnesses**: The agents can be used with Claude Code and other compatible harnesses, ensuring seamless integration into existing workflows. ### Tech Stack: - JavaScript ### Why It Matters: AI Employees offers a practical solution for businesses looking to automate their processes without the need for complex infrastructure. By providing pre-built agents and routines, it simplifies the implementation of AI solutions, making it accessible to a wide range of organizations. The agents can be easily integrated into existing workflows, reducing the need for custom development and speeding up the adoption process. ### Repository: The repository is maintained by Mark Fulton and is hosted on GitHub. It has received 764 stars, indicating high user interest and engagement. ### Primary Language: The primary language used in the project is JavaScript. ### GitHub Stars: The repository has been starred by 764 users, suggesting a significant amount of interest and adoption. ### Additional Resources: - **The Eight AI Employees**: A detailed overview of the agents and their roles.

Python 797 stars

nlp-uoregon/trankit

Trankit is a Light-Weight Transformer-based Python Toolkit for Multilingual Natural Language Processing

Overview: Trankit is a lightweight Transformer-based Python toolkit designed for multilingual natural language processing. It provides trainable pipelines for 100 languages and pre-trained pipelines for 56 languages, outperforming the state-of-the-art multilingual toolkit, Stanza, across 90 Universal Dependencies v2.5 treebanks of 56 different languages. Key features: - Trainable pipelines for 100 languages - Pre-trained pipelines for 56 languages - Auto Mode for multilingual pipelines - Command-line interface for easy usage Tech stack: Python Why it matters: Trankit offers a powerful and efficient solution for multilingual NLP tasks, making it suitable for general users and researchers alike.

Python 796 stars

Junchao-cs/SolarWM

Open data and scalable training for long-horizon video world models.

### Overview: SolarWM is a fully open-source framework designed for building diverse interactive video world models. It supports four 5B-33B models across Wan2.2, LTX-2.5, and MiniMax-H3, with a unified codebase that preserves each backbone's native representation and objective. ### Key Features: - **Four Models, One Codebase:** A unified framework supports four 5B-33B models across Wan2.2, LTX-2.5, and MiniMax-H3, with shared data, training, and inference interfaces. - **Simple Three-Stage Training Recipe:** Bidirectional adaptation, teacher-forced autoregressive initialization, and distribution matching distillation form a shared route across heterogeneous video backbones, without specialized ODE or consistency-distillation initialization. - **1.43M Clips · 25 TB. Fully Open:** We release the complete dataset and data-processing pipeline, unifying 14 datasets into a frame-aligned format with metric camera geometry, captions, quality metadata, and provenance. Source processing is decoupled from training-mixture design, allowing flexible data selection and mixing without reprocessing the source datasets. ### Tech Stack: Python ### Why It Matters: SolarWM provides a comprehensive solution for building interactive video world models, making it accessible and scalable for researchers and practitioners alike.

Jupyter Notebook 796 stars

MrGiovanni/ModelsGenesis

[MICCAI 2019 Young Scientist Award] [MedIA Best Paper Award] Models Genesis: self-supervised pre-training for 3D medical images. Learns transferable representations from unlabeled CT and MRI volumes, then fine-tunes for downstream segmentation and classification. Keras and PyTorch weights included.

### Overview: Models Genesis is a self-supervised learning framework designed for pre-training 3D medical images. It leverages non-linear intensity transformation, local pixel shuffling, out-painting, and in-painting to learn anatomy from unlabeled CT and MRI volumes. The pre-trained models can then be fine-tuned for downstream tasks such as segmentation and classification. ### Key Features: - **Self-Supervised Learning:** Models Genesis uses self-supervised learning techniques to pre-train models on unlabeled medical images. - **Transferable Representations:** It learns transferable visual representations that can be fine-tuned for various medical image analysis tasks. - **Foundation for Medical Imaging:** It serves as a foundation for medical imaging, particularly useful when labeled data is limited. - **Generative Family:** Models Genesis is part of the generative or restorative family of self-supervised learning methods, distinct from contrastive methods. ### Tech Stack: - **Primary Language:** Jupyter Notebook ### Why It Matters: - **Medical Image Analysis:** Models Genesis is a significant advancement in medical image analysis, providing a robust foundation for various medical imaging tasks. - **Foundation for Research:** It is a primary source of transfer learning for 3D medical imaging, making it invaluable for researchers and practitioners. - **Recognition Awards:** The repository has received prestigious awards, including the MICCAI 2019 Young Scientist Award and the MedIA Best Paper Award, highlighting its importance in the field. ### Repository: - **GitHub Stars:** 795 - **Description:** This repository provides the

Python 796 stars

samre12/deep-trading-agent

Deep Reinforcement Learning based Trading Agent for Bitcoin

### Overview: A deep reinforcement learning-based trading agent for Bitcoin, utilizing the DeepSense network for Q function approximation. This project leverages advanced reinforcement learning techniques to optimize trading strategies for Bitcoin prices. ### Key Features: - **Reinforcement Learning**: Utilizes Deep Reinforcement Learning (DRL) algorithms to make trading decisions. - **DeepSense Network**: Implements a neural network architecture inspired by the DeepSense paper for Q function approximation. - **Bitcoin Price Series**: Processes and analyzes Bitcoin price data to inform trading decisions. - **Training Parameters**: Supports parameter tuning through a configuration file for better performance. ### Tech Stack: - Python 2.7 - TensorFlow - Pandas - tqdm ### Why It Matters: This project aims to provide a robust and efficient trading solution for Bitcoin, leveraging advanced machine learning techniques to enhance investment strategies. ### Repository: - **Repository**: samre12/deep-trading-agent - **GitHub Stars**: 796 - **GitHub Description**: Deep Reinforcement Learning based Trading Agent for Bitcoin using DeepSense Network for Q function approximation. ### Requirements: - Python 2.7 - TensorFlow - Pandas - tqdm ### Docker Support: - **Docker Pull**: `docker pull samre12/deep-trading-agent:latest` - **Docker Run**: ```bash docker run -p 6006:6006 -it samre12/deep-trading-agent:latest ``` ### Support: - **Star the Repository**: Support the

Python 796 stars

timctho/convolutional-pose-machines-tensorflow

Overview: This repository contains the implementation of Convolutional Pose Machines (CPM) using TensorFlow, a state-of-the-art model for 2D body and hand pose estimation. Key features: - Easy multi-stage graph construction - Kalman filters for smooth pose estimation - Simple self-tracking module Tech stack: Python Why it matters: This project provides a practical and efficient solution for real-time 2D pose estimation, making it valuable for applications such as augmented reality, robotics, and gaming. Repository: timctho/convolutional-pose-machines-tensorflow Primary language: Python GitHub stars: 796 GitHub description: No description provided.

C++ 795 stars

OpenMOSS/MOSS-VL

An open-weight 11B model series for long-form and real-time video understanding

### Overview: MOSS-VL is an open-weight model series designed for long-form and real-time video understanding, featuring three models with 11B parameters each and open-source weights. It supports real-time interaction, offline use, and continuous dialogue. ### Key Features: - **MOSS-VL-Realtime**: Real-time interaction over continuous video streams, interruptible at any moment, answering on the fly, and deciding on its own when to respond and when to keep watching. - **MOSS-VL-Instruct**: Built for offline use, with particular strength in complex long-video understanding and in-depth dialogue. - **MOSS-VL-Base**: An open pre-trained foundation offering strong video–language representations for continued pre-training and downstream fine-tuning. ### Tech Stack: C++ ### Why It Matters: MOSS-VL enables real-time interaction on continuous video streams, achieving open-source SOTA latency in streaming interaction, and supports dynamic scene understanding and multi-turn dialogue.

Python 795 stars

ankandrew/fast-plate-ocr

Lightweight & fast OCR models for license plate text recognition.

### Overview: Fast and lightweight OCR models for license plate text recognition. Ideal for use after a plate object detector, these models are efficient and easy to integrate. ### Key features: - **Keras 3 Backend Support**: Train models using TensorFlow, JAX, or PyTorch backends. - **Efficient Execution**: Lightweight models that are cheap to run. - **ONNX Runtime Inference**: Fast and optimized inference with ONNX runtime. - **User-Friendly CLI**: Simplified CLI for training and validating OCR models. - **Region Recognition (Optional)**: Predict region/country of the license plate. - **Model HUB**: Access to pre-trained models ready for inference. - **Train/Fine-tune**: Easily train or fine-tune your own models. - **Export-Friendly**: Export models to CoreML, TFLite, or ONNX formats. ### Tech stack: Python ### Why it matters: These models are designed to be efficient and easy to integrate, making them ideal for use in applications requiring license plate recognition.

Rust 795 stars

bosun-ai/swiftide

Fast, streaming indexing, query, and agentic LLM applications in Rust

Overview: Swiftide is a Rust framework designed for building fast, streaming indexing, query, and aggregating language models (LLMs) applications. It provides an agent harness, typed task graphs for orchestration, and streaming RAG pipelines for efficient and composable large-scale data processing. Key features: - Agent harness for building autonomous agents - Typed task graphs for orchestrating tasks - Streaming RAG pipelines for large-scale indexing and retrieval Tech stack: Rust Why it matters: Swiftide simplifies the process of building LLM applications by providing a framework that abstracts away the complexity of handling LLM interactions, tools, and pipelines, making it easier to compose and orchestrate tasks in a composable manner. GitHub stars: 795 Repository: bosun-ai/swiftide Project Status: The project has been open for contribution and is actively maintained by the core team. Contributing: The core team welcomes contributions from the community to improve the framework and add new features. Core Team Members: The team includes individuals with expertise in Rust, LLMs, and large-scale data processing. License: The project is released under the MIT License, allowing for free use and modification.