Project directory

Projects

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

Unknown 284 stars

haofanwang/awesome-conditional-content-generation

Update-to-data resources for conditional content generation, including human motion generation, image or video generation and editing.

### Overview: This repository contains a collection of resources and papers on Conditional Content Generation, focusing on human motion generation, image generation, and video generation. It is maintained by Haofan Wang and is updated with the latest advancements in the field. ### Key Features: - **Human Motion Generation**: Includes papers on music-driven motion generation, text-driven motion generation, audio-driven motion generation, and motion prediction. - **Image Generation**: Covers text-image generation and text-3D image generation. - **Video Generation**: Focuses on text-video generation. ### Tech Stack: The repository utilizes diffusion models and other advanced techniques for generating conditional content. ### Why it Matters: This repository is essential for researchers and practitioners interested in the field of conditional content generation, providing a comprehensive overview of the latest developments and applications in the area.

Python 284 stars

lightonai/fast-plaid

High-Performance Engine for Multi-Vector Search

Overview: FastPlaid is a high-performance engine designed for multi-vector search, replacing single document or image vectors with per-token vectors to capture nuanced relevance accurately. Key features: - Multi-vector search - Late interaction mechanism for fine-grained similarity calculation - PLAID (Per-Token Late Interaction Dense Search) - Blazing performance optimized for GPUs Tech stack: Rust and Python Why it matters: FastPlaid enhances traditional vector search methods by capturing subtle relevance that single-vector models overlook, leading to higher accuracy and improved retrieval performance.

Python 284 stars

lzyhha/VisualCloze

[ICCV 2025] VisualCloze: A universal image generation framework that can support a wide range of in-domain tasks and generalize to unseen ones. (🔥 🔥 🔥 Merged into offical pipelines of diffusers.)

Overview: VisualCloze is a universal image generation framework that supports a wide range of in-domain tasks and can generalize to unseen ones through visual in-context learning. It has been merged into the official pipelines of diffusers and is used in the Hugging Face Gradio demo. Key features: 1. Support various in-domain tasks. 2. Generalize to unseen tasks through in-context learning. 3. Unify multiple tasks into one step and generate both target image and intermediate results. 4. Support reverse-engineering a set of conditions from a target image. Tech stack: Python Why it matters: VisualCloze is a powerful tool for image generation that can handle a variety of tasks and improve the model's ability to generalize to new situations, making it valuable for researchers and developers working in the field of computer vision and image processing. GitHub stars: 284

Python 284 stars

mamei16/LLM_Web_search

An extension for oobabooga/text-generation-webui that enables the LLM to search the web

### Overview: This project enables local LLMs to search the web, enhancing their capabilities with the ability to execute native tool calls or output commands for web searches. ### Key features: - **Web Search Integration**: Allows LLMs to perform web searches directly from the model's output. - **Command Execution**: Native tool calls and command output can be used to initiate web searches. - **Regular Expression Matching**: Web pages are extracted from search results using a specified regular expression. ### Tech stack: Python ### Why it matters: This extension enhances the functionality of local language models by providing them with the ability to perform web searches, which is crucial for applications requiring web-based information retrieval.

Jupyter Notebook 284 stars

mrm8488/shared_colab_notebooks

A Repo to store the Google Colaboratory Notebooks that I have created and shared

Overview: This repository stores and shares Google Colaboratory Notebooks created and modified by the user. It includes a variety of projects in Natural Language Processing (NLP), Computer Vision, Generative Adversarial Networks (GANs), and more. Key features: - NLP/NLG Colabs: Basic self attention, T5 base finetuned common_gen, T5 evaluate WMT, T5 finetuned wikiSQL demo, T5 qgsquad with HF transformers for QG, T5 wikiSQL multitask with HF transformers, DialoGPT using 🤗Transformers, disaster identification using tweeter colab, distilGPT2 finetuned MA Meditations, NER data augmentation, TF emotion detection, fine tuning on PGN English for NER, fine tuning on UDPOS English for POS, Fine tuning Transformers for typo detection, toxicity simple transformers. - Computer Vision Colabs: Pretrain ViT HF, Fine tune ConvNeXT with HF trainer beans dataset, Fine tune ConvNeXT with HF trainer, keras ocr custom, keras ocr custom v2, 3D Ken Burns, 3D Photo Inpainting, Deeper 3D Ken Burns single pic, custom learningtopaint, YOLACT++, TWINGAN, Face Depixelizer, FUNIT MonsterMirror v1, generate colors from text. - GANs/Misc: 3D Ken Burns, 3D Photo Inpainting multiple download, Deeper 3D Ken Burns single pic, custom learningtopaint, YOL

TypeScript 284 stars

xianjianlf2/MindGeniusAI

An AI agent that reads your PDFs and draws editable mind maps — visible tool-calling loop, built-in RAG, bring-your-own-key, multi-provider (OpenAI / Claude / DeepSeek / Kimi). Self-hostable.

### Overview: MindGenius AI is a self-hostable AI agent that converts PDFs into editable mind maps. It runs entirely on infrastructure controlled by the user, ensuring privacy and security. ### Key Features: - **Self-Hosted & Private**: One Docker image, no documents or keys leave your machine. - **Real Agent, Not a Prompt Wrapper**: Decides when to search documents, generate the map, and expand branches. - **Surgical Edits**: Edits nodes by id, preserving manual tweaks. - **Two-Way Canvas**: Edit nodes by hand and see the changes on Hermas's next message. - **Reads Your PDFs**: Retrieves PDFs from memory, allowing LangChain searches across attached documents. - **Keyboard-First Canvas**: Uses `Tab`, `Enter`, `F2`, and `Del` for navigation and editing. - **Import / Export**: Supports Markdown, OPML, PNG, SVG, Markdown, OPML, and Mermaid formats. - **Bring Your Own Key**: Allows using any compatible model endpoint. ### Tech Stack: TypeScript ### Why It Matters: MindGenius AI is designed for privacy-conscious knowledge workers, self-hosters, and teams with sensitive documents. It ensures your documents and keys never leave your machine, providing a secure and efficient way to create mind maps directly from your PDFs.

Python 283 stars

armbues/SiLLM

SiLLM simplifies the process of training and running Large Language Models (LLMs) on Apple Silicon by leveraging the MLX framework.

Overview: SiLLM simplifies the process of training and running Large Language Models (LLMs) on Apple Silicon by leveraging the MLX framework. It introduces additional features such as Low-rank Adaptation (LoRA) and Direct Preference Optimization (DPO) for enhanced LLM operations. Key features: - LLM Loading: Support for loading LLMs in different formats (Huggingface, Torch, GGUF, MLX) - LoRA Training: Train LLMs using Low-rank Adaptation - DPO Training: Train LLMs with Direct Preference Optimization - Experimental Features: Speculative Decoding, Beam Search, Logit Distillation, ... Tech stack: Python Why it matters: SiLLM aims to make large language model training and inference on Apple Silicon more accessible and efficient, leveraging the MLX framework to simplify the process and enhance performance. Repository: armbues/SiLLM GitHub stars: 282 GitHub description: SiLLM simplifies the process of training and running Large Language Models (LLMs) on Apple Silicon by leveraging the MLX framework. README excerpt: ![sillm](https://github.com/armbues/SiLLM/assets/4117144/859002e9-d209-480b-adb2-7276cd360cbe)

Jupyter Notebook 283 stars

naver/bergen

Benchmarking library for RAG

Overview: BERGEN is a benchmarking library designed to evaluate Retrieval-Augmented Generation (RAG) systems, focusing on question-answering tasks. It provides a framework for reproducible experiments and flexible configuration options. Key features: - Easy reproducibility and integration of new datasets and models - Support for various retrievers, rerankers, and large language models - Flexible YAML configuration system - Comprehensive evaluation metrics - Multilingual support Tech stack: Jupyter Notebook Why it matters: BERGEN helps researchers and developers benchmark and compare RAG systems effectively, ensuring consistent and reliable evaluations across different datasets and configurations. Repository: naver/bergen GitHub stars: 283 GitHub description: Benchmarking library for RAG README excerpt: [Image of BERGEN](documentation/images/BERGEN.png) Repository: naver/bergen GitHub stars: 283 GitHub description: Benchmarking library for RAG README excerpt: [Image of BERGEN](documentation/images/BERGEN.png) # Overview: BERGEN is a benchmarking library designed to evaluate Retrieval-Augmented Generation (RAG) systems, focusing on question-answering tasks. It provides a framework for reproducible experiments and flexible configuration options. Key features: BERGEN supports various retrievers, rerankers, and large language models, with a flexible YAML configuration system. It offers comprehensive evaluation metrics and multilingual support. Tech stack: BERGEN uses Jupyter Notebook as its primary technology stack. Why it matters: BERGEN helps researchers and

Python 283 stars

vakovalskii/searcharvester

Self-hosted search + markdown harvester for AI agents. SearXNG (100+ engines) + FastAPI + trafilatura. Tavily-compatible /search plus /extract with size presets and pagination. One-command Docker Compose.

### Overview: Searcharvester is a self-hosted search and extract tool designed for AI agents. It leverages SearXNG, FastAPI, and trafilatura for comprehensive search and markdown extraction. The web UI allows you to start, monitor, and manage research jobs, including a deep research agent team. ### Key features: - **Search**: Tavily-compatible search with over 100 engines, including images and videos. - **Extract**: URL → clean markdown with readability and Defuddle, size presets, and pagination. - **Deep Research**: A deep research agent team for comprehensive question answers. - **Web UI**: Live job management, model selection, and real-time monitoring. - **Security**: No API keys required, fully self-hosted, and uses GHCR for the adapter image. ### Tech stack: - Python ### Why it matters: - Self-hosted, secure, and customizable. - Comprehensive search and extract capabilities. - Real-time monitoring and management of research jobs. - Open-source and community-driven development. ### Repository: - **GitHub Repository**: [vakovalskii/searcharvester](https://github.com/vakovalskii/searcharvester) - **Stars**: 282 - **Description**: Self-hosted search + extract + deep research for AI agents, with a web UI to watch the agents work. Includes SearXNG, FastAPI, trafilatura, and Tavily compatibility. Supports size presets and pagination, and is fully self-hosted with a GHCR adapter image. ### Quick start

Python 282 stars

Cardinal-Operations/ORLM

ORLM: Training Large Language Models for Optimization Modeling

Overview: This project explores the training of open-source Large Language Models (LLMs) specifically for optimization modeling. ORLM aims to address the critical requirements for the training dataset of OR LLMs, design and implement OR-Instruct, and introduce the IndustryOR benchmark for testing LLMs on real-world optimization problems. Key features: - OR-Instruct: A semi-automated process for creating synthetic data tailored to specific optimization modeling requirements. - IndustryOR Benchmark: The first industrial benchmark for evaluating LLMs on solving real-world optimization problems. Tech stack: Python Why it matters: ORLM provides a significant improvement in the capability of optimization modeling by applying data from OR-Instruct to various open-source LLMs of 7B size, resulting in state-of-the-art performance on benchmarks like NL4OPT, MAMO, and IndustryOR. Repository: Cardinal-Operations/ORLM GitHub stars: 282 GitHub description: ORLM: Training Large Language Models for Optimization Modeling

Jupyter Notebook 282 stars

QuangBK/generativeAgent_LLM

Implementation of "Generative Agents: Interactive Simulacra of Human Behavior" paper with Guidance and Langchain. Full features and work with local LLMs.

### Overview: This project implements the "Generative Agents: Interactive Simulacra of Human Behavior" paper with guidance and Langchain, focusing on using local LLMs for improved performance and features like memory and dialogue generation. ### Key Features: - Work with local LLM - Memory and Retrieval - Reflection - Planning (improved) - Reacting and re-planning - Dialogue generation (improved) - Agent summary - Interview ### Tech Stack: Jupyter Notebook ### Why It Matters: This implementation aims to closely replicate the original paper's features, making it a valuable resource for developers interested in creating interactive simulations of human behavior using local language models.

Python 282 stars

Tencent-Hunyuan/ArtifactsBenchmark

### Overview: ArtifactsBench: Bridging the Visual-Interactive Gap in LLM Code Generation Evaluation ArtifactsBench is a new benchmark designed to evaluate the visual and interactive fidelity of Large Language Models (LLMs) in generating dynamic and interactive visual artifacts. This project aims to address the existing evaluation gap in the field, where established benchmarks focus on algorithmic correctness but lack consideration for visual fidelity and interactive integrity. ### Key Features: - **Benchmarking Tasks**: ArtifactsBench includes 1,825 diverse tasks, covering a wide range of use cases and scenarios. - **Evaluation Framework**: The framework programmatically renders each generated artifact and assesses it using an MLLM-as-Judge guided by a fine-grained, per-task checklist. - **Scoring Mechanism**: The evaluation is based on a holistic and reproducible scoring system, ensuring that the generated artifacts are not only correct but also visually and interactively aligned with human expectations. ### Tech Stack: - Python ### Why It Matters: ArtifactsBench provides a scalable and accurate tool for the community to accelerate the development of user-centric generative models. By addressing the evaluation gap, it helps researchers and practitioners focus on the critical aspects of visual and interactive fidelity, ultimately leading to more effective and engaging user experiences. ### GitHub Metadata: - Repository: **Tencent-Hunyuan/ArtifactsBenchmark** - Primary Language: **Python** - GitHub Stars: **282** - GitHub Description: **No description provided.** - README Excerpt: **No description provided.** ###

Python 282 stars

deep-symbolic-mathematics/LLM-SR

[ICLR 2025 Oral] This is the official repo for the paper "LLM-SR" on Scientific Equation Discovery and Symbolic Regression with Large Language Models

### Overview LLM-SR is a novel approach for scientific equation discovery and symbolic regression that leverages strengths of Large Language Models (LLMs). This method combines LLMs' scientific knowledge and code generation capabilities with evolutionary search to discover accurate and interpretable equations from data. The approach represents equations as program skeletons, allowing for flexible hypothesis generation guided by domain-specific priors. Experiments on custom benchmark problems across physics, biology, and materials science demonstrate superior performance compared to state-of-the-art symbolic regression methods, particularly in out-of-domain generalization. ### Key Features - **LLM-SR**: A novel method for scientific equation discovery and symbolic regression. - **Combination of LLMs' Knowledge and Code Generation**: Utilizes LLMs' scientific knowledge and code generation capabilities. - **Evolutionary Search**: Incorporates evolutionary search for flexible hypothesis generation. - **Program Skeleton Representation**: Equations are represented as program skeletons for flexible hypothesis generation. - **Superior Performance**: Demonstrates superior performance compared to state-of-the-art symbolic regression methods, especially in out-of-domain generalization. - **Challenging Datasets**: Proposed new, challenging datasets for evaluating LLM-based equation discovery methods. ### Tech Stack Python ### Why It Matters LLM-SR introduces a novel approach for scientific equation discovery and symbolic regression that leverages the strengths of large language models. This method demonstrates superior performance compared to state-of-the-art methods, particularly in out-of-domain generalization. The proposed datasets and benchmarks are designed to effectively test and evaluate LLM-based equation discovery methods.

Python 282 stars

pengfeng/ask.py

A simple Python program to implement the search-extract-summarize flow.

Overview: A simple Python program designed to implement the search-extract-summarize flow, similar to AI search engines. Key features: - Run locally with Ollama inference and embedding models. - Run on command line or with a GradIO UI. - Control output behavior, such as extracting structured data or changing output language. - Control search behavior, including restricting to a specific site or date, or scraping a list of URLs. - Automate complex search/data extraction tasks via cron jobs or bash scripts. - Ask questions against local files. Tech stack: Python Why it matters: This program is a demonstration of basic concepts in AI search engines, showcasing the raw constructs without focusing on performance or scalability. It provides a foundation for understanding the fundamentals of search and data extraction.

Python 282 stars

voicekit-team/T-one

T-one is a high-performance streaming ASR pipeline for Russian, specialized for the telephony domain.

Overview: T-One is a high-performance streaming Automatic Speech Recognition (ASR) pipeline designed specifically for the telephony domain, optimized for real-time transcription. Key features: - Pretrained CTC-based acoustic model - Streaming log-probabilities splitter - Modular components for easy integration - Real-time transcription using either greedy decoding or KenLM-based CTC beam search decoder Tech stack: Python Why it matters: T-One provides a practical and low-latency ASR solution tailored for the telephony industry, making it ideal for applications requiring real-time speech recognition in the field of telecommunications.

TypeScript 281 stars

BeatAPI/BeatDesign

Local-first AI canvas and timeline. Any MCP agent. Open-source alternative to Higgsfield.

### Overview: BeatDesign is a local-first AI canvas and timeline platform that allows users to build image and video workflows directly on their machine. It offers a seamless integration with MCP agents, enabling users to work on projects without the need for a cloud-based service. ### Key Features: - **Local Canvas and Editor:** Users can create and edit projects locally, ensuring no dependency on external servers. - **Project Settings:** Easily manage project settings, including renaming, selecting a cover image, and resetting to automatic or local video poster. - **Consistent Versions:** Canvas previews and Take selections agree on the latest successful output, preserving manual and concrete timeline assets. - **Faster Project Entry:** Choosing between Canvas or Studio creates a single project and opens it directly, with retry protection. - **Reliable Audio Export:** MCP timeline rendering preserves clip gain across sequential shots, with peak protection for overlapping audio. - **30 MCP Tools:** Agents can update project settings through the same revision-checked command path as the UI, ensuring compatibility and reliability. - **Browser-Native Export:** Supports browser-native MP4 export, eliminating the need for system dependencies. ### Tech Stack: TypeScript ### Why It Matters: BeatDesign provides a local-first alternative to Higgsfield, offering a connected creative workflow with the flexibility of tools like Higgsfield. It allows users to build and edit projects directly on their machines, ensuring no dependency on external servers. The platform supports 30 MCP tools and offers reliable audio export and consistent project versions, making it a valuable tool for AI image and video creators.

Unknown 281 stars

QiushiSun/Awesome-Code-Intelligence

Neural Code Intelligence Survey 2024-25; Reading lists and resources

Overview: This repository contains the survey paper "A Survey of Neural Code Intelligence: Paradigms, Advances and Beyond," authored by Qiushi Sun, Zhirui Chen, Fangzhi Xu, Kanzhi Cheng, Chang Ma, Zhangyue Yin, Jianing Wang, Chengcheng Han, Renyu Zhu, Shuai Yuan, Qipeng Guo, Xipeng Qiu, Pengcheng Yin, Xiaoli Li, Fei Yuan, Lingpeng Kong, Xiang Li, Zhiyong Wu, and others. Key features: - The survey covers various paradigms and advancements in Neural Code Intelligence. - It includes multiple reading lists and resources for further study. - It provides development timelines and timelines for the evolution of code intelligence. Tech stack: Unknown. Why it matters: This survey paper offers valuable insights into the current state and future directions of Neural Code Intelligence, providing a comprehensive overview of the field. Repository: QiushiSun/Awesome-Code-Intelligence GitHub stars: 281 GitHub description: Neural Code Intelligence Survey 2024-25; Reading lists and resources.

Python 281 stars

SkyworkAI/MoE-plus-plus

[ICLR 2025] MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts

### Overview: MoE++ accelerates Mixture-of-Experts methods with zero-computation experts, reducing computational overhead significantly. ### Key features: - **Zero-Computation Experts**: Discard, Skip, and Replace operations for zero-computation. - **Gating Residuals**: Enables each token to consider the previous layer's pathway. - **Efficient Inference**: Accelerates inference by minimizing computation. ### Tech stack: Python ### Why it matters: MoE++ offers a significant performance improvement with minimal computational overhead, making it ideal for real-time applications and resource-constrained environments.

Python 281 stars

binzhouchn/masr

中文语音识别系列，读者可以借助它快速训练属于自己的中文语音识别模型，或直接使用预训练模型测试效果。

Overview: MASR中文语音识别系列提供了一个快速训练和使用自己的中文语音识别模型的工具，适用于语音识别领域的开发人员和研究者。 Key features: - 开箱即用 - 自行训练 - 使用与训练分离（增量训练） - 高识别率 Tech stack: Python Why it matters: 该系列为开发者提供了一个便捷的工具，用于快速训练中文语音识别模型，或直接使用预训练模型进行测试，显著提高了语音识别的准确性和效率。 Repository: binzhouchn/masr Primary language: Python GitHub stars: 281 GitHub description: 中文语音识别系列，提供了一个快速训练和使用自己的中文语音识别模型的工具，适用于语音识别领域的开发人员和研究者。

Python 281 stars

fzliu/radient

Radient turns many data types (not just text) into vectors for similarity search, RAG, regression analysis, and more.

### Overview: Radient is a developer-friendly, lightweight library for generating and working with embeddings. It supports simple vectorization as well as building retrieval systems, making it useful for tasks like similarity search, RAG, and regression analysis. ### Key features: - **Simple Vectorization**: Radient allows you to easily convert text, audio, images, and more into vectors. - **End-to-End Workflow**: Radient provides a `Workflow` object to build complex vector-centric ETL applications. - **Performance Optimization**: Radient's `accelerate` function optimizes vectorizers for performance, reducing computation time. - **Lightweight Library**: Radient is lightweight and easy to integrate into Python projects. ### Tech stack: Radient uses the `sentence-transformers` library for text vectorization and supports other models like Mixbread AI's `mxbai-embed-large-v1`. ### Why it matters: Radient simplifies the process of working with embeddings, making it easier for developers to integrate embeddings into their projects without needing to dive deep into the underlying libraries. Its lightweight nature and performance optimization make it a valuable tool for anyone looking to leverage embeddings in their applications.