Projects

Overview: This repository is an open collection of methodologies designed to help LLM training engineers and operators in successfully training large language models. Key features: - Model parallelism - Maximizing throughput - Tensor precision / Data types - Training hyper-parameters and model initializations - Instabilities - Debugging software and hardware failures - SLURM - Resources Tech stack: Python Why it matters: This repository provides practical scripts and commands to quickly address common challenges in large language model training, making it a valuable resource for the LLM training community.

### Overview: MimirQ 是一个企业级的RAG（Reinforcement Learning with Attention）知识库系统，旨在通过可控的解析、治理、切块、混合检索、重排、引用、图谱、评测和Dify接入，实现企业知识的高效管理和应用。 ### Key features: - **可控解析**：从文档到证据的每一步都能检查输入、输出与版本。 - **可替换策略**：按文档和场景切换解析器、索引、检索与模型。 - **业务切块**：按标题、章节、业务记录或父子关系切块。 - **混合召回**：结合BM25、向量检索与重排。 - **重排与引用**：确保每一步都能检查输入输出、追溯引用与版本。 - **Golden回归**：用Golden题集守住发布质量。 ### Tech stack: Python ### Why it matters: MimirQ 是一个企业级的RAG知识库系统，通过可控的解析、治理、切块、混合检索、重排、引用、图谱、评测和Dify接入，实现企业知识的高效管理和应用。它解决了企业知识库交付中常见的问题，如错误可定位、策略可替换、质量可回归等，适用于各种复杂业务场景。

### Overview: This curated library contains 150 production prompts for OpenAI's GPT Image models, including GPT Image 2.5. Each prompt is reproduced verbatim from its original author, with their account and a link to the post. The prompts are grouped by use case rather than style, reflecting the demand's natural language shape. Every prompt retains its original wording, author, and source. ### Key Features: - 150 production prompts - 110 creators credited - 9 use cases - 25 written as JSON - Updated 2026-09-10 ### Tech Stack: JavaScript ### Why It Matters: These prompts are a curated collection of natural-language commands designed to generate images using OpenAI's GPT Image models. They are useful for anyone looking to create specific images using the models, without needing to run them on the latest version of GPT Image 2.5. The prompts are designed to be easily adaptable and can be used across different versions of the models. ### Repository: YouArt Open Source ### GitHub Stars: 591 ### GitHub Description: GPT Image 2.5 prompts, verbatim and credited: 150 community prompts across 9 use cases, each with the author who wrote it and their post. Free to run. ### Primary Language: JavaScript ### GitHub Actions: - [License](https://img.shields.io/badge/license-see%20LICENSE-blue.svg) - [PRs welcome](https://img.shields

Overview: This project is a diffusion model for generating novel protein backbone structures. It leverages trigonometry and attention mechanisms to create accurate protein structure predictions. Key features: - Diffusion model for protein backbone generation - Trigonometry and attention mechanisms - Available on HuggingFace Spaces and SuperBio Tech stack: Python, PyTorch, PyTorch Lightning, HuggingFace transformers Why it matters: This model offers a novel approach to protein structure prediction, potentially accelerating the development of new drugs and understanding protein functions.

### Overview: This repository is a survey paper and its associated project, focusing on AI-assisted and automated scientific research. It tracks papers across the full research lifecycle, covering stages such as Idea Generation, Literature Review, Coding & Experiments, Tables & Figures, Writing, Peer Review, Rebuttal & Revision, and Dissemination. The project aims to provide a roadmap and user guide for researchers interested in AI auto-research. ### Key features: - **Full Research Lifecycle Tracking**: Tracks papers from Idea Generation to Dissemination. - **AI-Assisted Phases**: Covers Peer Review, Rebuttal & Revision. - **Paper Documentation**: Includes a paper on AI auto-research and a project page with additional details. ### Tech stack: HTML ### Why it matters: This repository serves as a comprehensive resource for researchers interested in AI auto-research, providing a roadmap and user guide for the entire research lifecycle. It helps in understanding the current state of the field and can be a valuable tool for academic and professional development. ### Repository: worldbench/awesome-ai-auto-research GitHub stars: 565

### Overview: STCN is a novel, effective, and efficient framework designed specifically for video object segmentation. It achieves state-of-the-art results while running at high speeds, making it suitable for real-time applications. The key contributions include an improved memory coverage method and the use of L2 similarity for better performance and memory efficiency. ### Key features: - **Space-Time Correspondence Networks (STCN)**: A framework for modeling space-time correspondences in video object segmentation. - **L2 Similarity**: A method to replace dot product, significantly improving memory bank utilization. - **Improved Memory Coverage**: A technique to enhance memory usage and reduce computation time. - **High Performance**: Runs at 20+ FPS without bells and whistles, and even faster with mixed precision. ### Tech stack: Python ### Why it matters: STCN stands out by providing a simple yet effective solution for video object segmentation, achieving high accuracy and efficiency. Its simplicity makes it suitable for real-time applications, and the use of L2 similarity ensures better performance and memory efficiency compared to traditional methods.

Overview: Firefox Translations is a webextension that enables client side translations for web browsers, providing a user-friendly interface for translating web pages. Key features: - Client-side translations - Installation on Mozilla Add-ons - Supported languages: Spanish, Estonian, English, German, Czech, Bulgarian, Portuguese, Italian, French, Polish, Russian, Persian (Farsi), Icelandic, Norwegian Nynorsk, Norwegian Bokmål, Ukrainian, and Dutch Tech stack: The extension is developed using JavaScript, HTML, and CSS. Why it matters: Firefox Translations simplifies the process of translating web pages, making it easier for users to access content in their preferred language, enhancing the overall user experience. Repository: mozilla/firefox-translations GitHub stars: 567 GitHub description: Firefox Translations is a webextension that enables client side translations for web browsers. Firefox Translations was developed with The Bergamot Project Consortium, coordinated by the University of Edinburgh with partners Charles University in Prague, the University of Sheffield, University of Tartu, and Mozilla. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 825303. 🇪🇺 Primary language: JavaScript GitHub description: Firefox Translations is a webextension that enables client side translations for web browsers. Firefox Translations was developed with The Bergamot Project Consortium, coordinated by the University of Edinburgh with partners Charles University in Prague, the University of Sheffield, University of Tartu, and Mozilla. This project has

Overview: TabDDPM is an innovative method for modeling tabular data using diffusion models, presented in the ICML 2023 paper "TabDDPM: Modelling Tabular Data with Diffusion Models". This implementation is the official code for the paper's findings. Key features: - Implementation of TabDDPM model - Tuned hyperparameters for evaluation models (CatBoost, MLP) - Detailed documentation and examples for reproducibility Tech stack: Python Why it matters: TabDDPM offers a novel approach to modeling tabular data using diffusion models, providing a significant advancement in the field of tabular data analysis. Repository: yandex-research/tab-ddpm GitHub stars: 567

Overview: This repository is a curated list of resources for AGI Agents, including articles, videos, papers, and projects. It aims to provide a comprehensive overview of the latest advancements in AI agents and their applications. Key features: - Articles and videos - Papers - 2024-2026 important projects - Frontline projects - Agents development platforms - Curated list - Star history Tech stack: Python, JavaScript, and various machine learning frameworks. Why it matters: This repository serves as a valuable resource for developers and researchers interested in AGI agents, offering insights into the latest developments and practical applications in the field. Repository: yzfly/Awesome-AGI-Agents Primary language: Unknown GitHub stars: 567 GitHub description: 🤖 Awesome list of AGI Agents. Agents 精选资源合集. README excerpt: # Awesome-AGI-Agents Agents (智能体) 精选资源合集，持续更新中 ## 目录 - [文章和视频](#文章和视频) - [论文](#论文) - [2024-2026 重要项目](#2024-2026-重要项目) - [前沿项目](#前沿项目) - [Agents 开发平台](#agents-开发平台) - [Curated List](#curated-list) - [Star History](#star-history) ## 文章和视频 |名称|简介|备注|

Overview: LinearRAG is a relation-free graph construction method designed for efficient GraphRAG, which significantly reduces LLM token costs during graph construction, making it faster and more efficient than traditional methods. Key features: - Relation-free graph construction - Lightweight entity recognition and semantic linking - Multi-hop reasoning in a single retrieval pass without explicit relational graphs - Zero LLM token consumption, faster processing speed, and linear time/space complexity Tech stack: Python Why it matters: LinearRAG offers a significant improvement in performance by eliminating the need for LLM token costs during graph construction, making it more efficient and scalable for large-scale corpora. Repository: DEEP-PolyU/LinearRAG Primary language: Python GitHub stars: 564 GitHub description: [ICLR 2026] LinearRAG: Linear Graph Retrieval-Augmented Generation on Large-scale Corpora README excerpt: # **LinearRAG: Linear Graph Retrieval-Augmented Generation on Large-scale Corpora** > A relation-free graph construction method for efficient GraphRAG. It eliminates LLM token costs during graph construction, making GraphRAG faster and more efficient than ever. <p align="center"> <a href="https://arxiv.org/abs/2510.10114" target="_blank"> <img src="https://img.shields.io/badge/Paper-Arxiv-red?logo=arxiv&style=flat-square" alt="arXiv:2506.08

Overview: This repository contains a collection of projects written in various programming languages, suitable for beginners and advanced developers alike. The goal is to provide a centralized resource for finding projects, especially for beginners who might struggle to find projects on their own. Key features: - Beginner-level projects - Intermediate-level projects - Advanced-level projects Tech stack: Jupyter Notebook Why it matters: This repository serves as a valuable resource for developers, offering projects across different levels of expertise. It simplifies the process of finding projects, especially for beginners, by providing a centralized location.

### Overview: Multi-Modality Arena is an evaluation platform designed to benchmark large multi-modality models, providing a comprehensive evaluation of vision-language models. ### Key features: - OmniMedVQA dataset: A large-scale comprehensive evaluation benchmark for medical large-scale vision-language models. - Tiny LVLM-eHub: An evaluation benchmark for large vision-language models, featuring 4 models and 42 text-related visual benchmarks. - LVLM-eHub: An evaluation benchmark for publicly available large multimodal models, evaluating 8 models across 6 categories with 47 datasets and an online platform. ### Tech stack: Python ### Why it matters: This platform provides a standardized way to evaluate and compare vision-language models, ensuring fair and comprehensive assessments across different datasets and models.

### Overview: This repository provides fast inference solutions for the BLOOM model, specifically the 176B version. It supports both Hugging Face's accelerate and DeepSpeed Inference frameworks for generation tasks. The scripts are tested on 8 A100 GPUs for BLOOM 176B (fp16/bf16) and 4 A100 GPUs for BLOOM 176B (int8). The scripts are designed to be efficient and can be used for a variety of applications. ### Key features: - **Hugging Face Accelerate**: Supports generation tasks. - **DeepSpeed Inference**: Supports generation tasks. - **Model Compatibility**: Supports BLOOM 176B (fp16/bf16) and BLOOM 176B (int8). - **Deployment**: Uses logic borrowed from DeepSpeed MII library. - **Model Name**: Supports both `bigscience/bloom` and `microsoft/bloom-deepspeed-inference-fp16` for different deployment frameworks. ### Tech stack: Python ### Why it matters: This repository provides a fast and efficient way to perform inference tasks on the BLOOM model, making it suitable for applications that require quick response times and high throughput. The use of DeepSpeed Inference and Hugging Face's accelerate frameworks ensures that the models can be deployed efficiently on modern GPU setups, making it a valuable resource for developers working with large language models.

Overview: Omnichar is an open-source project that allows developers to create and manage AI characters using a single .char file. This file contains all the necessary elements such as face, body, wardrobe, and voice, ensuring consistency across different models and platforms. Key features: - Consistent character across different models and platforms - Generate locally on your own GPU, in ComfyUI, or with hosted models - Supports multiple models including FLUX.2, FLUX.1, Krea 2, Z-Image Turbo, MiniMax H3, and hosted models Tech stack: Python Why it matters: Omnichar simplifies the creation and management of AI characters by providing a single file that can be used across various models and platforms, ensuring consistency and ease of use.

### Overview: This is the Java language implementation of LangChain, designed to make it easy to develop LLM-powered applications. It bridges the realm of LLM within the Big Data domain, focusing primarily on the Java stack. ### Key features: - **LLMs**: Empowers LLMs with LangChain, supporting popular models like OpenAI, Azure OpenAI, ChatGLM2, and Ollama. - **Vector Stores**: Supports integration with vector stores like Pinecone and Milvus for efficient data retrieval. - **Big Data Integration**: Employs LangChain to build applications that can handle large datasets and complex queries. ### Tech stack: Java ### Why it matters: This project is essential for developers looking to integrate LLMs with Big Data environments, making it easier to build applications that leverage cutting-edge AI capabilities.

### Overview: TravelPlanner is a benchmark designed to evaluate language agents in planning tasks involving multiple constraints, including transportation, daily meals, attractions, and accommodation. It includes three types of constraints: Environment Constraint, Commonsense Constraint, and Hard Constraint. ### Key Features: - **Multiple Constraints Handling**: The benchmark evaluates language agents across various constraints to ensure robust performance. - **Comprehensive Planning**: Language agents must create detailed plans for each day, including transportation, meals, attractions, and accommodation. - **Real-World Relevance**: The constraints are based on real-world applications, making the benchmark highly relevant for practical use. ### Tech Stack: - **Programming Language**: Python - **Model**: GPT-4, LLMs, LLaMA-Factory, and other language models are used for fine-tuning and evaluation. ### Why It Matters: TravelPlanner provides a comprehensive benchmark for evaluating language agents in complex planning tasks, ensuring that agents can handle real-world constraints effectively. This benchmark is crucial for researchers and developers looking to improve the performance of language agents in planning applications. ### Repository: - **Repository**: OSU-NLP-Group/TravelPlanner - **GitHub Stars**: 563 - **GitHub Description**: "[ICML'24 Spotlight] 'TravelPlanner: A Benchmark for Real-World Planning with Language Agents'" ### Updates: - **2024/10/23**: Models finetuned on TravelPlanner, data available [here](./finetuning_data). - **2024/

Overview: This repository provides a simple example implementation of the VoiceRAG pattern, utilizing Azure AI Search and Azure OpenAI's gpt-4o-realtime-preview model to power interactive voice generative AI experiences. The pattern leverages RAG (Retrieval Augmented Generation) to integrate voice interactions with knowledge-based responses. Key features: - Voice interface for capturing user input - RAG (Retrieval Augmented Generation) to retrieve relevant documents and generate responses - Audio output to play the generated response Tech stack: Python Why it matters: This example showcases how to integrate voice-based AI experiences with Azure AI Search and Azure OpenAI's real-time API, making it a valuable resource for developers looking to create conversational AI applications that can understand and respond to voice inputs. Repository: Azure-Samples/aisearch-openai-rag-audio GitHub stars: 564

Overview: This project compiles papers and codes related to Large Language Models (LLMs) and Foundation Models (FMs) specifically designed for Time Series (TS) tasks. It aims to provide insights into these models for understanding their capabilities and applications. Key features: - Papers and codes for LLMs and FMs for TS tasks - Papers and codes for surveying large models for TS and ST data - Papers and codes for foundation models for TS analysis - Papers and codes for time series forecasting with LLMs - Papers and codes for time series forecasting with FMs - Papers and codes for time series forecasting with small models Tech stack: Python, PyTorch, TensorFlow, and other relevant tools for data processing and model training. Why it matters: This project serves as a comprehensive resource for developers and researchers interested in leveraging large language and foundation models for time series forecasting and analysis, providing a solid foundation for understanding and implementing these models effectively.

### Overview: TextRL is a thin, opinionated layer built on top of HuggingFace's TRL that simplifies the implementation of reinforcement learning for text generation. It provides a single configuration class, multiple trainer classes for various algorithms, and supports callable reward functions. This makes it easier to integrate modern text-generation RL into projects. ### Key Features: - **Configuration**: A single configuration class for all algorithms. - **Trainers**: Multiple trainer classes for different algorithm families. - **Callable Rewards**: Support for callable reward functions. - **Support for Preference and Reward Models**: Algorithms like DPO, IPO, and KTO can be trained with preference and reward models. - **Legacy API Removal**: The legacy PFRL/gym API is deprecated and replaced with unified `loss_type` for preference algorithms. ### Tech Stack: - Python ### Why It Matters: TextRL simplifies the implementation of reinforcement learning for text generation, making it easier to integrate modern RL techniques into projects. It supports various algorithms and reward models, making it a versatile tool for researchers and developers working on text generation tasks.

Overview: This project provides a Unity SDK for IBM Watson services, allowing developers to integrate Watson-powered applications into their Unity projects. Key features: - IBM Watson SDK for Unity - Tone Analyzer API - Natural Language Classifier API Tech stack: C# Why it matters: This SDK simplifies the integration of Watson services into Unity projects, making it easier for developers to create applications that utilize IBM Watson's AI capabilities.