Project directory

Projects

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

Jupyter Notebook 369 stars

JeongJiHeon/ScoreDiffusionModel

The Pytorch Tutorial of Score-based and Diffusion Model

### Overview: Explore the tutorial and blog posts on Score-based and Diffusion Models using PyTorch. Organized in Korean, these resources will help you understand the models better by looking at the code alongside the text. ### Key features: - Pytorch Tutorial of Score-based and Diffusion Model - Organized in Korean - Code and blog posts - Includes tutorials for NCSN, DDPM, DDIM, VESDE/VPSDE, and LDM - Improves techniques for training Score-Based Generative Models - Improves Denoising Diffusion Probabilistic Models - Diffusion Models Beat GANs on Image Synthesis - CLASSIFIER-FREE DIFFUSION GUIDANCE ### Tech stack: PyTorch ### Why it matters: These tutorials and blog posts provide a comprehensive guide to Score-based and Diffusion Models, making it easier to understand and implement these models in your projects.

Python 369 stars

MLT-OSS/open-assistant-api

The Open Assistant API is a ready-to-use, open-source, self-hosted agent/gpts orchestration creation framework, supporting customized extensions for LLM, RAG, function call, and tools capabilities. It also supports seamless integration with the openai/langchain sdk.

### Overview: Open Assistant API is an open-source, self-hosted AI intelligent assistant API that supports various extensions for LLM, RAG, function call, and tool capabilities. It integrates seamlessly with the official OpenAI SDK and offers a local deployment option. ### Key features: - **Custom Extensions**: Supports customized extensions for LLM, RAG, function call, and tool capabilities. - **RAG Engine**: Compatible with the R2R RAG engine, offering preliminary implementation for better RAG capabilities. - **One API Integration**: Supports integration with the One API for seamless integration with commercial and private models. - **Code Interpreter**: Under development, offering a code interpreter feature. - **Multimodal Support**: Supports multimodal capabilities. - **LLM Support**: Supports more LLM models, including GPT. - **Message Streaming Output**: Supports message streaming output for a smoother user experience. - **Local Deployment**: Supports local deployment and expansion of existing features. ### Tech stack: Python ### Why it matters: Open Assistant API is an open-source, self-hosted AI intelligent assistant API that supports various extensions for LLM, RAG, function call, and tool capabilities. It integrates seamlessly with the official OpenAI SDK and offers a local deployment option.

C# 369 stars

SciSharp/SiaNet

An easy to use C# deep learning library with CUDA/OpenCL support

Overview: SiaNet is an easy-to-use C# deep learning library that supports multiple backends including CNTK, TensorFlow, MxNet, PyTorch, and ArrayFire. It offers CUDA/OpenCL support for some backends, making it suitable for developers looking to create and train deep neural network models. Key features: - Multiple backend support - CUDA/OpenCL support for some backends - Light-weight library built with .NET Standard 2.0 - Easy to use library, focusing on research - Code well structured, easy to extend with new layers, loss, metrics, optimizers, constraints, and regularizers Tech stack: C# Why it matters: SiaNet simplifies the development of deep learning models in C#, making it accessible to developers without extensive deep learning experience. Its support for multiple backends and GPU acceleration through CUDA/OpenCL opens up a wide range of possibilities for research and practical applications.

Python 369 stars

Socialpranker/deepdive

Deepdive skill for Claude Code — 12-phase research pipeline: plan-review gate, parallel sub-agent search, claims-ledger triangulation with dissent protection, relevance × authority evidence filter, multi-angle red team, four-layer citation verification. 105 blocks, 29 channels, 460+ stat sources, 47 APIs, 1072 verified endpoints.

### Overview: Deepdive is a structured research skill designed for Claude Code, providing a 13-phase pipeline for investigating topics. This skill includes hypothesis testing, parallel sub-agent search, source triangulation, and adversarial review, ensuring every claim is traceable and reusable. ### Key Features: - **12-Phase Research Pipeline**: Plan-Review, Parallel Sub-Agent Search, Claims-Ledger Triangulation, Relevance × Authority Evidence Filter, Multi-Angle Red Team, Four-Layer Citation Verification. - **105 Blocks, 29 Channels, 460+ Stat Sources, 47 APIs, 1072 Verified Endpoints**: Comprehensive and detailed research methodology. - **100% API-Validation**: All endpoints are auto-validated weekly. ### Tech Stack: - Python ### Why It Matters: Deepdive ensures that every claim is traceable and reusable, making the research process efficient and transparent. The structured approach helps in avoiding ad-hoc research and ensures that the results are robust and reliable. ### Repository: - Repository: [Socialpranker/deepdive](https://github.com/Socialpranker/deepdive) - GitHub Stars: 368 - GitHub Description: A structured research skill for Claude Code with 12-phase pipeline, 105 blocks, 29 channels, 460+ stat sources, 47 APIs, and 1072 verified endpoints. ### Installation: - **For Claude Code (CLI)**: Clone the repository and

Python 369 stars

VinAIResearch/LFM

Official PyTorch implementation of the paper: Flow Matching in Latent Space

### Overview: Flow matching in latent space is an innovative framework that leverages flow matching techniques within the latent spaces of pretrained autoencoders to improve the efficiency and scalability of high-resolution image synthesis. This approach addresses the computational challenges associated with traditional flow matching methods by reducing the need for expensive pixel space evaluations. ### Key Features: - **Latent Space Flow Matching**: Utilizes pretrained autoencoders to match the latent flows of the generated images to the true data distribution, enhancing computational efficiency and scalability. - **Conditional Generation**: Explores label-conditioned image generation, image inpainting, and semantic-to-image generation, demonstrating effectiveness across various datasets. - **Theoretical Control**: Provides a theoretical control over the Wasserstein-2 distance between the reconstructed latent flow distribution and the true data distribution, ensuring upper-bounded performance. ### Tech Stack: Python 3.10 and PyTorch 1.13.1/2.0.0 ### Why It Matters: This method stands as a significant contribution in the integration of various conditions into flow matching, offering improved quality and flexibility for conditional generation tasks. By reducing the computational burden, it enables flow-matching training on constrained resources, making it particularly useful for real-time applications and resource-limited environments. --- Repository: VinAIResearch/LFM Primary language: Python GitHub stars: 368 GitHub description: Official PyTorch implementation of the paper: Flow Matching in Latent Space

Python 369 stars

diambra/arena

DIAMBRA Arena: a New Reinforcement Learning Platform for Research and Experimentation

### Overview DIAMBRA Arena is a comprehensive software package designed for Reinforcement Learning research and experimentation. It offers a collection of high-quality environments for various arcade emulated video games, providing a Python API that is fully compliant with OpenAI Gym/Gymnasium format. The package supports all major operating systems (Linux, Windows, and macOS) and can be easily installed via Python PIP. It is completely free to use, and users can register on the official website to access the platform. ### Key Features - **High-Quality Environments**: Games such as Space Invaders, Pac-Man, and others. - **Python API**: Fully compliant with OpenAI Gym/Gymnasium format. - **Comprehensive Documentation**: A detailed guide for easy setup and usage. - **Standard Interface**: Supports both single player (1P) and two players (2P) modes. - **Operating System Compatibility**: Works on Linux, Windows, and macOS. - **Installation**: Easy installation via Python PIP. - **Compliance with Standards**: Adheres to OpenAI Gym/Gymnasium format. - **Comprehensive Examples**: Includes examples covering various use cases, making it easy to start experimenting. ### Tech Stack - Python ### Why It Matters DIAMBRA Arena provides a standard interface for popular video games, making it an ideal resource for researchers and experimenters looking to explore various Reinforcement Learning subfields. Its compatibility with multiple operating systems and ease of installation make it accessible to a wide audience. ### GitHub Repository - Repository: [diambra/

Python 369 stars

krea-ai/flux-krea

Official GitHub repository for FLUX.1 Krea [dev].

Overview: This repository contains the official code and weights for the FLUX.1 Krea model, developed in collaboration with Black Forest Labs. The model offers superior aesthetic control and image quality, and can be used for image generation through inference code and a Jupyter Notebook. Key features: - Inference code - Jupyter Notebook for model execution - Model weights hosted on Huggingface Tech stack: Python Why it matters: The FLUX.1 Krea model is a significant contribution to the field of image generation, offering high-quality and aesthetic control, and is open-sourced to promote collaboration and innovation in AI. Repository: krea-ai/flux-krea GitHub stars: 369 GitHub description: Official GitHub repository for FLUX.1 Krea [dev].

Python 369 stars

lucidrains/deformable-attention

Implementation of Deformable Attention in Pytorch from the paper "Vision Transformer with Deformable Attention"

### Overview: This repository implements Deformable Attention, an improved version of Deformable Attention from the paper "Vision Transformer with Deformable Attention" by Zhuofan Xia et al. It is a PyTorch implementation that addresses issues with DETR and improves positional embedding extrapolation. ### Key features: - Deformable Attention implementation. - Improved relative positional embedding for better extrapolation. - Supports 1D, 2D, and 3D deformable attention. - Compatible with PyTorch. ### Tech stack: Python ### Why it matters: This implementation provides a more efficient and effective alternative to existing methods, offering improved performance and better handling of spatial attention in vision transformers.

Python 369 stars

robustsam/RobustSAM

RobustSAM: Segment Anything Robustly on Degraded Images (CVPR 2024 Highlight)

Overview: RobustSAM is a refined version of the Segment Anything Model (SAM) designed specifically for robust image segmentation on degraded images. It addresses the limitations of SAM by enhancing its performance on low-quality images while maintaining its flexibility and zero-shot capabilities. Key features: - RobustSAM enhances SAM's performance on degraded images. - It uses pre-trained SAM model with minimal parameter changes. - RobustSAM can be optimized within 30 hours on eight GPUs, making it practical for research labs. - Robust-Seg dataset is introduced for optimal training and evaluation. Tech stack: Python Why it matters: RobustSAM offers a significant improvement in image segmentation performance on degraded images, making it valuable for applications requiring robust zero-shot capabilities.

Unknown 369 stars

ymcui/PERT

PERT: Pre-training BERT with Permuted Language Model

### Overview: PERT is a pre-trained BERT model that leverages permuted language model to learn text semantics without using masked tokens. ### Key features: - **BERT Pre-training**: Perturbs the input text to form permuted sequences, which is then used to train a pre-trained BERT model. - **No Mask Tokens**: Perturbs the input text without introducing masked tokens, ensuring that the model can learn from the original token positions. - **Performance on Specific Tasks**: PERT shows improved performance on some NLU tasks but performs poorly on others. - **Model Sizes**: Available in two sizes: base (110M parameters) and large (330M parameters). ### Tech stack: Python, PyTorch, TensorFlow ### Why it matters: PERT provides an alternative approach to pre-training BERT models by leveraging permuted language models, offering improved performance on specific NLU tasks while maintaining a more lightweight model size compared to traditional masked token-based methods.

Unknown 368 stars

Allenpandas/Reinforcement-Learning-Papers

📚 List of Top-tier Conference Papers on Reinforcement Learning (RL)，including: NeurIPS, ICML, AAAI, IJCAI, AAMAS, ICLR, ICRA, etc.

### Overview: This repository is dedicated to showcasing the most significant research papers in the field of Reinforcement Learning (RL) that have been accepted at top academic conferences such as AAAI, IJCAI, NeurIPS, ICML, ICLR, ICRA, and AAMAS. It serves as a convenient resource hub to help developers stay updated on the latest developments in reinforcement learning, explore research trends, and delve into cutting-edge algorithms and methods. ### Key Features: - **Repository**: Allenpandas/Reinforcement-Learning-Papers - **GitHub Stars**: 368 - **Tech Stack**: None specified - **Why it Matters**: This repository provides a curated list of top-tier conference papers on reinforcement learning, making it an essential resource for researchers and developers interested in the latest advancements in the field. ### Table of Contents: - **Refine by Year**: Lists papers by year, including 2023, 2022, 2021, 2020, 2019, 2018, and 2017. - **Refine by Theme**: Categories include Multi-Agent Reinforcement Learning, Meta Reinforcement Learning, Hierarchical Reinforcement Learning, Multi-Task Reinforcement Learning, Offline Reinforcement Learning, Inverse Reinforcement Learning, and Reinforcement Learning with Large Language Models. - **1 Multi-Agent Reinforcement Learning**: Papers focusing on the optimization of multi-step predictions in model-based multi-agent reinforcement learning. - **2 Meta Reinforcement Learning**:

Unknown 368 stars

HKUDS/Awesome-LLM4Graph-Papers

[KDD'2024] "LLM4Graph: A Survey of Large Language Models for Graphs"

Overview: This repository serves as a collection of recent advancements in employing large language models (LLMs) for modeling graph-structured data. It categorizes and summarizes approaches based on four primary paradigms and nine secondary-level categories. Key features: - GNNs as Prefix - LLMs as Prefix - LLMs-Graphs Integration - LLMs-Only Tech stack: Unknown Why it matters: This repository provides valuable insights into the latest research in using LLMs for graph learning tasks, offering a comprehensive overview of various methodologies and their applications.

Unknown 368 stars

Jinfeng-Xu/Awesome-Multimodal-Recommender-Systems

[TMM'26] Continuously Updated Awesome Multimodal Recommendation Paper List

Overview: This repository is a survey on multimodal recommender systems, categorizing existing models into four critical aspects: Feature Extraction, Encoder, Multimodal Fusion, and Loss Function. It includes a list of papers on multimodal recommender systems and a comprehensive benchmark and code base for multimodal recommendations, the MRLib. Key features: - A survey on multimodal recommender systems - A list of papers on multimodal recommender systems - A continuously updated library for advanced models in multimodal recommendation Tech stack: Unknown Why it matters: This repository provides a valuable resource for researchers and developers interested in multimodal recommender systems, offering insights into the latest advancements and future directions in the field.

Python 368 stars

alipay/PainlessInferenceAcceleration

Accelerate inference without tears

### Overview: PIA, or Painless Inference Acceleration, is a toolkit designed to accelerate Large Language Model (LLM) inference without the hassle of tears. It includes three key features: `FLOOD`, `LOOKAHEAD`, and `IPAD`. ### Key Features: - **FLOOD**: A pure pipeline parallelism framework that enhances inference throughput by reducing communication costs, making it the successor to `LOOKAHEAD` for optimal performance across both small and large batch sizes. - **LOOKAHEAD**: Uses an on-the-fly trie-tree cache to prepare hierarchical multi-branch drafts, reducing the need for speculative decoding or additional head training, significantly improving generated token count in forward passes. - **IPAD**: Applies iterative pruning and distillation techniques to reduce model size, contributing to more efficient inference. ### Tech Stack: Python ### Why It Matters: PIA simplifies the inference process, making it more efficient and less resource-intensive, ensuring faster and more accurate model outputs.

TypeScript 368 stars

feyninc/chonkiejs

🦛 CHONK your texts with Chonkie ✨ Type-friendly, light-weight, fast and super-simple chunking library

Overview: CHONK your texts with Chonkie!✨ The no-nonsense lightweight and efficient chunking library. Key features: - Fast text chunking - Type-friendly and TypeScript-compatible - Ideal for RAG applications - No bloat, just CHONK Tech stack: TypeScript Why it matters: 🚀 Feature-rich, ✨ Easy to use, ⚡ Fast, 🪶 Light-weight, 🦛 Cute CHONK mascot, ❤️ [Moto Moto]'s favorite TypeScript library GitHub stars: 368 Repository: feyninc/chonkiejs

Python 368 stars

goruck/home-generative-agent

AI agent for Home Assistant — talk to your home, create automations in plain language, analyze cameras with face recognition, and get proactive anomaly alerts. Cloud LLMs or fully local.

Overview: Home Generative Agent is an AI agent for Home Assistant that enables conversational control over smart home entities, automation creation in plain language, camera analysis with face recognition, and proactive anomaly alerts. Key features: - Conversational control over smart home entities - Automation creation in plain English - Camera & image analysis with face recognition - Proactive anomaly alerts with Sentinel engine Tech stack: Python Why it matters: Home Generative Agent integrates seamlessly with Home Assistant, providing a conversational interface and automation capabilities, while also offering advanced security features like anomaly detection and face recognition. It's designed to be fully local or cloud-based, ensuring privacy and security.

Jupyter Notebook 368 stars

tirthajyoti/Deep-learning-with-Python

Deep learning codes and projects using Python

Overview: This repository contains a collection of deep learning code examples, tutorial-style Jupyter notebooks, and projects, all built on Python and utilizing various libraries such as TensorFlow, Keras, and Google Colab. Key features: - A variety of deep learning code examples - Tutorial-style Jupyter notebooks - Projects based on Python and various libraries - Utilizes Google Colab for some notebooks, requiring specific functions exclusive to Google Colab - Includes a utility module for example notebooks - A general-purpose regression module for tabular datasets Tech stack: Python, NumPy, Pandas, Matplotlib, TensorFlow, Keras, Google Colab Why it matters: This repository serves as a comprehensive resource for learning and practicing deep learning using Python, providing practical examples and projects that can be easily replicated and extended. Repository: tirthajyoti/Deep-learning-with-Python GitHub stars: 368 GitHub description: Deep learning codes and projects using Python

Jupyter Notebook 368 stars

wshuyi/demo-chinese-text-binary-classification-with-bert

### Overview: BERT模型在中文文本二元分类任务中的应用，通过快速、简便的Python代码实现，无需复杂的数据处理流程和大量的参数调整，使得普通用户也能轻松上手。 ### Key features: - 使用Google Colab Notebook提供了一个简明易懂的BERT模型应用示例。 - 代码简洁，仅需几行即可完成文本分类任务，无需额外的数据处理和参数调整。 - 适用于自然语言文本分类，包括情感分类等任务。 - 教程简单易懂，适合Python初学者和快速上手的用户。 ### Tech stack: Google Colab Notebook ### Why it matters: BERT模型在自然语言处理任务中表现出色，而Google Colab Notebook为普通用户提供了易于理解和使用的工具，使得BERT模型的使用门槛大大降低。

Python 367 stars

MME-Benchmarks/Video-MME-v2

Video-MME-v2: Towards the Next Stage in Benchmarks for Comprehensive Video Understanding

### Overview: Video-MME-v2 is a comprehensive benchmark designed to evaluate the capabilities of video understanding models. It builds upon the success of the Video-MME benchmark, which has been widely used for evaluating Gemini and GPT models. The new version, Video-MME-v2, aims to address the saturation of scores on existing benchmarks by introducing a progressive and robust evaluation system. This system includes three levels of evaluation: multi-point information aggregation, temporal understanding, and complex temporal reasoning. The evaluation mechanism is non-linear, focusing on both individual accuracy and the consistency and completeness of reasoning. ### Key Features: 1. **Progressive Multi-Level Evaluation Dimensions**: The benchmark decomposes video understanding into three levels—finding information, modeling time, and cross-temporal reasoning. 2. **Grouped Non-Linear Evaluation Mechanism**: Questions are organized into groups targeting capability consistency and reasoning coherence, with a non-linear scoring scheme that considers both individual accuracy and overall consistency. 3. **Rigorous Data Annotation**: The dataset is meticulously annotated to ensure high-quality evaluation. 4. **Leaderboard Performance**: The benchmark aims to provide a more accurate reflection of model capabilities compared to existing benchmarks. ### Tech Stack: Python ### Why It Matters: Video-MME-v2 is crucial for driving the next generation of video understanding models. By introducing a more robust and progressive evaluation system, it helps in identifying and addressing the gaps in current evaluation paradigms, ensuring that models are evaluated based on their true capabilities rather than just their scores on existing benchmarks. This will lead to higher-quality technical iteration and better understanding of

Python 367 stars

Yarroudh/segment-lidar

Python package for segmenting LiDAR data using Segment-Anything Model (SAM) from Meta AI.

### Overview: `segment-lidar` is a Python package designed for instance segmentation of aerial LiDAR data using the Segment Anything Model 3 (SAM 3) from Meta AI. This model is particularly useful for extracting objects from LiDAR point clouds, making it ideal for applications in geomatics and environmental monitoring. ### Key Features: - **Automatic Mask Generation**: SAM 3 automatically generates instance masks for objects in the LiDAR point cloud. - **Prompted Segmentation**: Users can provide specific prompts (e.g., "buildings", "trees") to enhance the segmentation accuracy. - **3D-to-2D Mapping**: The package projects the 3D point cloud into an image and then uses SAM 3 to find instances. - **Multiple Outputs**: Supports automatic mask generation, text prompts, box prompts, top views, pinhole cameras, saved label rasters, ground filtering, and LAS or LAZ output formats. - **Model Dependency**: Requires the `facebook/sam3` model from Hugging Face, which can be downloaded locally for inference. ### Tech Stack: The primary tech stack used in `segment-lidar` includes Python, SAM 3, and the Hugging Face model repository. ### Why It Matters: This package is essential for applications in geomatics, environmental monitoring, and urban planning, where accurate segmentation of LiDAR data is crucial. It simplifies the process of extracting meaningful information from LiDAR scans, making it easier to analyze and visualize the data in a more intuitive way. ### Repository: The repository for