Project directory

Projects

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

Jupyter Notebook 16 stars

ssG12333/AI-Infra-Inference-Notes

AI Infra 推理系统学习资料：从深度学习算子到 vLLM 服务端部署

Overview: 这个项目提供了一套系统化的自学资料，旨在帮助开发者从深度学习算子学习到 vLLM 服务端部署的推理系统全面掌握。资料涵盖了从基础算子到高级架构，再到性能优化和部署实战的多个模块。 Key features: - 系统架构笔记 - 模块化学习资料 - 自检测试 - 学习计划 - 简历模板 - 面试问答 - 术语速查 Tech stack: Jupyter Notebook Why it matters: 本项目为 AI Infra 推理系统的学习提供了全面的资料和体系，适合希望从深度学习算法转向推理工程或大模型部署的开发者，以及对系统级知识感兴趣的读者。

Astro 16 stars

sugarcane-ai/sugarcane-ai.github.io

### Overview: Sugarcane AI: Open Source Microservices Framework for LLM App Development 🤖 Sugarcane AI is an open-source framework designed to simplify and accelerate the development of language model (LLM) applications. With a focus on fine-tuned language models, prompt management, and workflow plugins, Sugarcane AI empowers developers to build, train, and manage complex LLM applications effortlessly. 🎉 ### Key Features: - **Simplify Workflow Development**: Develop, train, and manage workflow plugins seamlessly, reducing complexity over the chain of prompts. - **Reusable Prompt Packages**: Create, share, and reuse high-accuracy prompts with ease, saving time and effort, and facilitating rapid experimentation in production without regression, ensuring performance visibility at every step. - **Open Source Community**: Collaborate and contribute to the thriving open-source community of developers dedicated to simplifying AI app development using Sugarcane AI. ### Tech Stack: Astro ### Why it Matters: Simplifying AI App Development Sugarcane AI’s microservices architecture for LLM applications offers a game-changing approach to AI app development, bringing flexibility, scalability, reliability, and efficiency compared to monolithic LLM application structures. ### Getting Started - **Install Sugarcane AI**: Follow the installation instructions provided in the README. - **Explore Documentation**: Dive into the comprehensive documentation for a detailed guide on how to use Sugarcane AI effectively. - **Contribute**: Join the community and contribute to the open-source project to help improve the framework. ### Demo Videos - **Introduction

R 16 stars

tejas4patel/dragonfarm

Raise your own dragon: fine-tune a language model on your own documents, from R. Writes the training data, trains on your laptop, RAG built in.

Overview: Dragonfarm is a tool designed to train a language model directly from your own documents using R, without the need for Python or API fees. It includes a RAG (Reinforcement Learning with Agents) component and allows for fine-tuning and monitoring of the model's performance. Key features: - Fine-tuning a language model directly from your documents. - Built-in RAG (Reinforcement Learning with Agents) for answers with sources. - Training on your own GPU or a free cloud GPU. - Fine-tuning speeds up to 1.5 to 2.8 times faster than the standard trainer. - Comes with a draft GxP validation pack for regulated teams. - Built-in unsloth for fine-tuning. - Compatible with ellmer and MCP. Tech stack: R Why it matters: Dragonfarm simplifies the process of training a language model, making it accessible without the need for complex programming languages or expensive APIs. It allows users to train models directly from their documents, improving efficiency and reducing costs.

Python 16 stars

tien02/ensemble-roberta-fasttext-vietnamese

Ensemble PhoBERT with FastText Embedding to improve performance on Vietnamese Sentiment Analysis tasks.

Overview: This project aims to improve Vietnamese sentiment analysis performance by combining PhoBERT with FastText embeddings. The dataset used is the UIT-VSFC corpus, which contains over 16,000 sentences with balanced labels for sentiment analysis. Key features: - Ensemble PhoBERT with FastText embeddings - Optimized code directory structure - Containerized with Docker - Experimented with PhoBERT and FastText models Tech stack: Python Why it matters: This project demonstrates effective ensemble techniques for improving sentiment analysis models, particularly addressing the class imbalance issue in Vietnamese datasets.

JavaScript 16 stars

upgundecha/applied-ai

A repository of curated use cases, articles, blogs, videos on how companies are using Artificial Intelligence and Machine Learning.

Overview: This repository is a curated collection of artificial intelligence and machine learning use cases, best practices, and tools adopted by leading technology and tech-savvy organizations. It serves as a knowledge repository for topics such as AI, ML, DL, Generative AI, Large Language Models, Prompt Engineering, Retrieval-Augmented Generation, MLOps, Data Engineering, and Responsible AI. Key features: - Artificial Intelligence - Machine Learning - Deep Learning - Generative AI - Large Language Models - Prompt Engineering - Retrieval-Augmented Generation (RAG) - MLOps - Data Engineering - Responsible AI Tech stack: JavaScript Why it matters: This repository provides valuable insights and practices for developers and engineers looking to implement AI and ML solutions in their projects, contributing to the development of responsible and effective engineering cultures.

Python 16 stars

vbzdz/seedance-2-0-prompts-pt

Dreamina — Seedance 2.0 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedance-2.0`) and the cost per image.

### Overview: Seedance 2.0 Prompts: 8 Tested Recipes with Outputs Dreamina — Seedance 2.0 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedance-2.0`) and the cost per image. This repository provides 8 tested recipes along with their example renders, durations, and the cost per image on the Seedance 2.0 model. ### Key Features: - **Model ID:** `seedance-2.0` - **Aspect Ratios:** 16:9, 4:3, 1:1 - **Cost per Image:** $0.066 per second (480P), $0.216 per second (720P), $0.3849 per second (1080P) ### Tech Stack: - Python ### Why It Matters: Seedance 2.0 provides a cost-effective way to generate high-quality images without the need for a monthly subscription. The API is compatible with OpenAI and offers a simple pricing model, making it ideal for developers and content creators looking to streamline their workflow. ### Repository: - Repository: `vbzdz/seedance-2-0-prompts-pt` - GitHub Stars: 11 - GitHub Description: Dreamina — Seedance 2.0 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedance-2.0`) and the cost per image. This repository provides 8 tested recipes

HTML 16 stars

xianyu110/awesome-gpt-image2.5

精选 ChatGPT Images 2.5 / GPT-Image-2.5（Flare · Sunburst · Sketch）好玩用法画廊、可抄 prompt 与评测 · 在线试用 gptimage2.asia

### Overview: 这是一个精选的ChatGPT图片2.5/GPT-Image-2.5（Flare · Sunburst · Sketch）画廊，包含各种用法和提示词，旨在帮助开发者快速生成高质量的图像。该画廊提供免费在线试用，包括多种模板和可复制的提示词，适用于各种用途，如营销素材、个人作品展示等。 ### Key features: - **在线试用**：每条案例都有“一键试用”功能，可直接在gptimage2.asia生成器中生成提示词。 - **可复用工作流**：提供七个可复制的工作流技能，适用于各种场景，如控制形状、选择型、官方提示与选型、写真锁脸、局部编辑、像素精灵表→GIF、虚拟IP资产流水线等。 - **模板与提示词**：包括Flare、Sunburst、Sketch等多种模板和提示词，适用于各种用途，如海报排版、电商改图、场景视觉、像素动效、UX产品等。 ### Tech stack: - HTML ### Why it matters: 这个画廊为开发者提供了丰富的图像生成工具和提示词库，帮助他们在各种应用场景中快速生成高质量的图像，节省时间和精力。

Python 16 stars

yeungtuzi/Xiaotu-LLM

XTU = X Transformers Unity — hybrid (CPU-expert + GPU) inference acceleration for vLLM: mainline plugin + CPU MoE engine (AVX512-VNNI/BF16 MXFP4/FP8), long-prefill GPU streaming, SM80 (A100) port.

Overview: Xiaotu-LLM 是一个为超大 MoE 模型设计的推理架构，旨在通过 vLLM 插件形式提供高性能的推理能力。它支持在单张显卡且内存充足的设备上运行 DeepSeek-V4.1-Flash、GLM-5.3-Flash、MiMo-V2.6-Flash-RL 和 Qwen3.8-Flash-Next 等 MoE 模型。此外，它还支持自动 CPU/GPU Prefill 分流、支持 SM80 等显卡，以及硬件感知的 NUMA 切片等特性。 Key features: - 支持在单张显卡且内存充足的设备上运行超大 MoE 模型 - 自动 CPU/GPU Prefill 分流 - 支持 SM80 等较老的显卡 - 支持多种 ISA 的 CPU 计算后端 - 支持 GPU 常驻热专家和按策略执行动态换入 - 支持 GPU + CPU 协同解码，提升 DeepSeek-V4.1-Flash 的解码性能 - 支持 12 层精选专家层常驻，见 v0.2.9 发行说明 Tech stack: Python Why it matters: Xiaotu-LLM 是一个高性能的推理架构，适用于大规模 MoE 模型的推理任务，能够显著提升解码性能并减少显存占用，特别是在硬件

Python 16 stars

yeungtuzi/vllm-xtu-moe

XTU = X Transformers Unity — hybrid (CPU-expert + GPU) inference acceleration for vLLM: mainline plugin + CPU MoE engine (AVX512-VNNI/BF16 MXFP4/FP8), long-prefill GPU streaming, SM80 (A100) port.

Overview: Xiaotu-LLM 是一个面向超大 MoE 模型的推理架构，提供 vLLM 插件形式。它支持在单张显卡且内存充足的设备上运行 DeepSeek-V4.1-Flash、GLM-5.3-Flash、Qwen3.8-Flash-Next 等 MoE 模型，通过自动 CPU/GPU Prefill 分流、支持 NVFP4 / MXFP4 量化权重、硬件感知的 NUMA 切片等特性，显著提升了解码性能。 Key features: - 支持在单张显卡且内存充足的设备上运行超大 MoE 模型 - 自动 CPU/GPU Prefill 分流 - 支持 NVFP4 / MXFP4 量化权重 - 硬件感知的 NUMA 切片 - 支持 GPU 常驻热专家，并按策略执行动态换入 - 支持 GPU + CPU 协同解码，显著提升 DeepSeek-V4.1-Flash 的解码性能 - 支持 12 层精选专家层常驻，见 v0.2.9 发行说明 Tech stack: Python Why it matters: Xiaotu-LLM 提供了高效、稳定的推理架构，适用于大规模 MoE 模型的推理，特别是在资源有限的设备上，显著提升了解码性能和模型的运行效率。

Unknown 16 stars

ysh329/awesome-deep-learning-finetune-experience

This repository not only contains experience about parameter finetune, but also other in-practice experience such as model ensemble (boosting, bagging and stacking) in Kaggle or other competitions.

Overview: This repository is a collection of experiences and practical insights related to deep learning finetuning and ensemble methods, including model ensemble techniques such as boosting, bagging, and stacking, and various competitions on Kaggle. Key features: - Model ensemble methods (boosting, bagging, stacking) - Kaggle competitions experience - Deep learning finetuning techniques - Hyper-parameter tuning and optimization - Generative Adversarial Networks (GANs) Tech stack: Python, Scikit-learn, TensorFlow, PyTorch, Keras Why it matters: This repository provides valuable insights and practical examples for developers interested in improving their deep learning models through finetuning and ensemble methods, as well as for those looking to enhance their Kaggle competition performance through model ensemble techniques.

Python 16 stars

z0ydfhoij/wan2-7-image-prompts-id

Wan 2.7 Image prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`wan2.7-image`) and the cost per image.

### Overview: This repository contains tested image prompt recipes for the Wan 2.7 model, including 8 recipes with examples, aspect ratios, and the cost per image. ### Key Features: - **8 Recipes**: Eight distinct image prompts for the Wan 2.7 model. - **Aspect Ratios**: The aspect ratios for each recipe. - **Cost per Image**: The cost of each image generated on this route. ### Tech Stack: Python ### Why it Matters: This repository provides a concise guide for developers and designers to create effective image prompts using the Wan 2.7 model, ensuring high-quality and cost-effective results.

Python 16 stars

zhangjiayang6835-cyber/ai-research

🍯 自进化 AI 能力研究平台：蜜罐系统 → 沙箱评测引擎 → 训练数据集与 LoRA 微调流水线 / Self-Evolving AI Security Research Platform

Overview: 自进化 AI 能力研究平台是一个完整的 AI 安全研究闭环系统，通过蜜罐系统发布安全任务，沙箱评测引擎多维度评估，最终将结果沉淀为训练数据反哺模型微调。 Key features: - 蜜罐系统发布安全任务，捕获 AI 代码修复行为 - 沙箱评测引擎多维度评估，包括 Docker 沙箱执行、6 种作弊检测、评测指标报告生成 - 训练场提供标准数据集格式和 LoRA 微调流水线 Tech stack: Python Why it matters: 该平台为 AI 安全研究提供了完整的闭环系统，有助于提高 AI 模型的安全性，并通过实验数据反哺模型微调，进一步提升模型的性能和安全性。 Repository: zhangjiayang6835-cyber/ai-research Primary language: Python GitHub stars: 15 GitHub description: 🍯 自进化 AI 能力研究平台：蜜罐系统 → 沙箱评测引擎 → 训练数据集与 LoRA 微调流水线 / Self-Evolving AI Security Research Platform README excerpt:  ![Tasks](https://img.shields.io/badge/tasks-0-blue) ![Contributors](https://img.shields.io/badge/contributors-15-brightgreen) ![Bounty](https://img.shields.io/badge/bounty-$4770-gold) ![Submissions](

Python 16 stars

zhengyima/Anchors

Source code of CIKM2021 Paper 'Pre-training for Ad-hoc Retrieval: Hyperlink is Also You Need'

Overview: This project is a source code repository for the CIKM 2021 paper "Pre-training for Ad-hoc Retrieval: Hyperlink is Also You Need". It includes pre-training on a corpus based on hyperlinks and fine-tuning on MS MARCO Document Ranking datasets. Key features: - Pre-training on corpus based on hyperlinks - Fine-tuning on MS MARCO Document Ranking datasets - Preinstallation guide for Python3 environment - Pre-training corpus data preparation - Shell script for generating pre-training samples - Pre-training script for generating pre-training data Tech stack: Python Why it matters: This project provides a comprehensive solution for pre-training and fine-tuning models for ad-hoc retrieval tasks, demonstrating the effectiveness of pre-training on hyperlink data in improving retrieval performance.

Python 15 stars

404stillhere/upg

Universal Prompt Generator — prompt-engineering system for text-to-image models. Spectrum of 1–5 calibrated prompts per theme, with safety/drift/cliche enforcement built into the pipeline.

Overview: UPG is a prompt-engineering system designed for text-to-image models, offering a calibrated spectrum of 1-5 prompts per theme. It includes safety, drift, and cliché enforcement to ensure high-quality outputs. Key features: - Six-phase pipeline with explicit validation gates - Safety transform layer to enforce safety, drift, and cliché - Theme-width aware enrichment for broad and specific themes - EIS-based drift control - BAN REGISTRY to prevent diluting prompts - Slot system S1-5 with Ring Test - Eight Creative Lenses (PHOTO, CINE, TEXTURAL, PAINT, ENV, SYMBOLIC, MINIMAL, SENSORY) Tech stack: Python Why it matters: UPG provides a robust and customizable prompt system for text-to-image models, ensuring high-quality and safe outputs while maintaining the integrity of themes and avoiding clichés or diluting the intended message. Repository: 404stillhere/upg Primary language: Python GitHub stars: 15 GitHub description: Universal Prompt Generator — prompt-engineering system for text-to-image models. Spectrum of 1–5 calibrated prompts per theme, with safety/drift/cliche enforcement built into the pipeline. README excerpt: A prompt-engineering system for text-to-image models. Takes a theme, returns a calibrated spectrum of 1–5 executable prompts, with safety, drift, and cliché enforcement built into the pipeline.

Python 15 stars

4m4n5/CLIP-Lite

Pytorch Implementation of CLIP-Lite | Accepted at AISTATS 2023

Overview: CLIP-Lite is a PyTorch implementation of a lightweight version of CLIP, accepted at AISTATS 2023. This project is designed for efficient downstream tasks such as classification and detection. Key features: - PyTorch implementation - CLIP-Lite architecture - Compatible with Detectron2 for pretraining - Supports COCO, PASCAL VOC, ImageNet, and iNaturalist datasets - Pre-processes COCO Captions into LMDB files for faster data reading Tech stack: Python Why it matters: CLIP-Lite offers a lightweight solution for downstream tasks, making it suitable for applications where computational resources are limited. Its compatibility with popular frameworks like Detectron2 ensures seamless integration for pre-training and fine-tuning.

Python 15 stars

AI-Partner-Cool/SimpleClassification

Image Classification Tutorial: ConvNext--> 98.8% on CIFAR10 + 92.4% on CIFAR100; ResNet18 -- 95.6% on CIFAR10 + 79.1% on CIFAR100

Overview: This repository contains an image classification tutorial using ConvNext and ResNet18 models, achieving high accuracy on CIFAR10 and CIFAR100 datasets. It also includes pretraining results using Inet1K and MocoV2. Key features: - Supports ConvNext and ResNet18 models - Provides pretraining results with Inet1K and MocoV2 - Supports CIFAR10, CIFAR100, CUB, and CARS datasets - Uses Tensorboard for visualization and EMA tracking Tech stack: Python Why it matters: This tutorial offers a practical guide for developers to enhance image classification models, demonstrating significant improvements in accuracy and reliability metrics. Repository: AI-Partner-Cool/SimpleClassification GitHub stars: 15

Jupyter Notebook 15 stars

Abeshith/RAG-FundaMentals

🔰 A Comprehensive RAG repository covering basic vanilla RAG techniques, advanced retrieval methods, hybrid search fusion approaches, hands-on reranking techniques with code + explanation 📚✨

### Overview: This repository provides a comprehensive guide to RAG implementations, covering various techniques from basic to advanced. It includes practical notebooks demonstrating different RAG approaches, retrieval strategies, and optimization methods. ### Key Features: - **Basic RAG**: Core pipeline with vector embeddings, semantic search, and prompt engineering. - **Document RAG**: Processes PDF documents for local knowledge base and metadata handling. - **URL RAG**: Real-time retrieval for web content, with updates to dynamic knowledge. - **Contextual Compression**: Compresses retrieved documents to extract only relevant portions. - **Hypothetical Document Embedding (HyDE)**: Generates hypothetical answers to improve retrieval accuracy. - **Multi-Hop Retrieval**: Step-by-step retrieval for complex queries. - **Parent Document Retrieval**: Retrieves small chunks but returns larger parent documents. - **Self-Query Retrieval**: Natural language query parsing with metadata filtering. - **Sentence Window Retrieval**: Expanded context windows around relevant sentences. ### Tech Stack: Jupyter Notebook ### Why It Matters: This repository offers hands-on experience with various RAG techniques, making it valuable for developers looking to implement and optimize RAG systems.

Jupyter Notebook 15 stars

Anshumaan-Chauhan02/Guided-Flow-Matching

Utilized attention incorporated UNet model for conditional image generation using Flow Matching with Conditional Optimal Transport Objective

### Overview: The project integrates Conditional Optimal Transport into an attention-based UNet model for both conditional and unconditional image generation tasks. Utilizing a Classifier Free Guidance (CFG) mechanism, it addresses the descriptive limitations of the CIFAR10 dataset, enhancing the conditioning process with the BLIP2 FLAN T5 model. The self and cross attention mechanism, incorporating timestep and tokenized text, facilitates conditioning. Extensive experimental analysis leads to an optimized architecture with a FID score of 105.54 for unconditional generation and CLIPScore/FID scores of 22.19/305.42 for conditional generation. The research highlights the model's potential, suggesting further improvements through architectural refinements and extended training. ### Key Features: - Conditional image generation - Conditional flow matching - Classifier Free Guidance (CFG) - Attention-based UNet model - Conditional and unconditional image generation - Optimized architecture with high FID scores ### Tech Stack: Jupyter Notebook ### Why it Matters: The model demonstrates significant potential in conditional image generation, offering improvements in FID scores and CLIPScore/FID scores compared to existing methods. This makes it a valuable tool for applications requiring precise image generation based on conditional inputs.

C# 15 stars

AutomationAgent-AI/GenAI.NET

Implements a framework to build Generative AI applications.

### Overview: GenAI.NET is a .NET library designed to interact with Large Language Models, including OpenAI, providing features such as a language model, function calling, generative AI service, conversation interface, function toolset, DLL function toolset, prompt template, pipeline, map-reduce tool, search tool, and agents. ### Key features: - **Language Model:** Connects to OpenAI to retrieve responses. - **Function Calling:** Allows passing function descriptions to the language model. - **GenerativeAI Service:** Provides high-level functions. - **Conversation:** Retains conversation chains and uses vector stores for context. - **Function Toolset:** Executes functions from a toolset. - **DLLFunctionToolset:** Wraps public static methods of a .NET class. - **PromptTemplate:** Parametersize prompts with template strings. - **Pipeline:** Sequentially executes tools. - **MapReduceTool:** Processes large data sets in parallel. - **SearchTool:** Performs web searches using Bing API. - **Agents:** Plans and executes actions based on available tools. ### Tech stack: C# ### Why it matters: GenAI.NET simplifies the process of building Generative AI applications, making it easier for developers to integrate and utilize Large Language Models in their projects.

Python 15 stars

BoHuangLab/CELL-E_2

Multimodal encoder-only transformer model for image-based protein predictions

### Overview: CELL-E 2 is a multimodal encoder-only transformer model designed for image-based protein predictions. This project aims to translate protein sequences into images and vice versa, providing a bidirectional approach for protein structure and function analysis. ### Key features: - Image-based protein prediction and generation - Sequence prediction using a transformer model - Supports both image and sequence inputs - Utilizes a bidirectional text-to-image transformer architecture ### Tech stack: Python ### Why it matters: CELL-E 2 offers a novel method for analyzing protein structures and functions by leveraging multimodal data, making it valuable for researchers and bioinformaticians seeking to understand protein interactions and dynamics.