Projects

Overview: VortexKV is a high-performance, pure Go in-memory data engine designed for developers seeking a fast, scalable, and secure alternative to Redis. It offers unparalleled throughput and low latency, with peak throughput of 14,287,238 ops/sec and a latency of 111 µs. Key features: - **Redis Protocol Compatibility**: Fully compatible with Redis, allowing seamless integration with existing Redis-based applications. - **World-Record Throughput**: Peak throughput of 14,287,238 ops/sec (GET P=128), 10,000,000 ops/sec (GET P=64), and 2,941,490 ops/sec (SET P=128). - **Multi-Reactor Engine**: Custom event-driven reactor architecture for efficient CPU core pinning and elimination of CPU migration and cache invalidation. - **Smart Socket Pipeline Coalescing**: Batches pipelined responses into consolidated kernel writes, reducing syscall context-switching by 95%. - **Cacheline-Padded 256-Shard Concurrency**: Eliminates CPU L1/L2 false sharing across cores and removes global client mutex bottlenecks. - **Enterprise Production Security**: Full `requirepass` and `AUTH [username] <password>` support, native TLS/SSL wire encryption, and web studio security shield modal for HTTP endpoints and WebSockets. - **Resource Safety & MaxMemory LRU Eviction**: Strict memory cap with active `

Overview: MultimodalHugs is a lightweight framework that simplifies the training, evaluation, and deployment of multimodal AI models using minimal code differences from Hugging Face, ensuring seamless compatibility with their pipelines. Key features: - Minimal boilerplate - Reproducible pipelines - Modular design - Built to work with Hugging Face native - Examples included Tech stack: Python Why it matters: MultimodalHugs streamlines the development process by providing a standardized setup and minimal code differences, making it easier to experiment with multimodal AI models without sacrificing performance or compatibility. GitHub stars: 19

### Overview: FaceXBench is a comprehensive benchmark designed to evaluate Multimodal Large Language Models (MLLMs) on complex face understanding tasks. It includes 5,000 multimodal multiple-choice questions derived from 25 public datasets and a newly created dataset, FaceXAPI. The benchmark covers 14 tasks across 6 key categories, assessing MLLMs' face understanding abilities in bias and fairness, face authentication, recognition, analysis, localization, and tool retrieval. ### Key Features: - **14 Tasks Across 6 Categories:** Assess MLLMs' face understanding abilities in various aspects. - **5, 000 Questions:** Covering 14 tasks across 6 broad categories. - **Strategic Design:** Questions include four options, with only one correct answer, prompting models to analyze carefully. - **Benchmark for Evaluation:** Introduces a comprehensive benchmark for evaluating MLLMs on face understanding tasks. ### Tech Stack: - Python ### Why It Matters: FaceXBench is a crucial resource for developing MLLMs capable of performing sophisticated face understanding, addressing the significant room for improvement in current models.

Overview: 多模态股价预测系统 TCNT 是一个结合了多种数据模态（如文本、图像、音频等）的股票预测模型，旨在提高股价预测的准确性。 Key features: - 多模态股票数据集 - 预训练模型Chinese-RoBERTa-wwm-ext - 实验结果展示 Tech stack: Python Why it matters: 通过整合多种数据模态，多模态股价预测系统 TCNT 提高了股价预测的准确性和鲁棒性，为投资者提供了更全面的分析工具。 Repository: Lanping-Tech/Multi-modal-Valuation-Forecast-System Primary language: Python GitHub stars: 19 GitHub description: 多模态股价预测系统 [百度云](https://pan.baidu.com/s/1pOHj0VViGEd2ECN8kg0mcA) 提取码: kq84

Overview: nab is a token-lean web microfetcher that converts any URL into clean markdown. It supports real browser cookies, on-device multilingual ASR, and MCP protocol for authenticated access. The output is optimized for language models. Key features: - Token-lean fetching - Multilingual ASR and OCR - Real browser cookie support - MCP protocol authentication - URL watching and notifications - Prompt-injection defense by default Tech stack: Rust Why it matters: nab provides a fast, secure, and efficient way to fetch and analyze web content, making it ideal for language models and other applications requiring clean, context-aware data. GitHub stars: 19 Repository: MikkoParkkola/nab

Overview: This project is a web application designed to automate the process of news scraping and summarization, allowing users to input article URLs and receive concise summaries of the news articles. It supports both local inference with a fine-tuned BART model and remote deployment via FastAPI or Streamlit interfaces. Key features: - Automated News Scraping: Fetches news articles from specified URLs. - Content Summarization: Utilizes a fine-tuned BART model for summarization. - User Interface: Offers a Streamlit application for local development and a deployed Streamlit frontend for remote access. - Configurable: Supports local configuration through a `config.json` file. Tech stack: Python Why it matters: This application streamlines the news reading experience by automatically summarizing articles, making it convenient for busy users and enhancing the overall user experience. Repository: NhanPhamThanh-IT/Deepnews-Summarizer Primary language: Python GitHub stars: 19

Overview: Kaizen is an implementation of practical self-supervised continual learning with continual fine-tuning, showcasing significant improvements in accuracy and flexibility in the face of continual learning challenges. Key features: - Kaizen introduces a training architecture that mitigates catastrophic forgetting for both feature extractor and classifier. - It uses a carefully designed loss function to balance knowledge retention and learning from new data. - Demonstrated up to 16.5% accuracy improvement on split CIFAR-100 benchmarks compared to previous SSL models. Tech stack: Python Why it matters: Kaizen provides a practical solution for continual learning, offering flexibility and improved performance in computer vision tasks, making it valuable for applications requiring incremental learning. Repository: Nokia-Bell-Labs/kaizen GitHub stars: 19 GitHub description: Official implementation of the algorithm outlined in "Kaizen: Practical Self-Supervised Continual Learning With Continual Fine-Tuning" by Chi Ian Tang et al.

Overview: This project provides a zero-dependency Python SDK for interacting with WeChat Bot (iLink protocol) for text, image, and video messaging, as well as a MCP (Message Center Protocol) server. Key features: - Zero-dependency Python SDK - Supports text, image, and video messaging - MCP server functionality Tech stack: Python Why it matters: This project simplifies the integration of WeChat Bot functionality into Python applications, making it easier for developers to build applications that can send and receive messages through the iLink protocol and interact with a MCP server.

Overview: A batch-1 decode engine for Bonsai 27B (ternary and 1-bit GGUF packs) on NVIDIA GPUs, replacing the vendor's llama.cpp fork for token generation and producing the same outputs, up to 1.53x faster on H100 compared to the vendor's published numbers. Key features: - Faster token generation with up to 1.53x performance gain on H100 compared to the vendor's fork. - Single-launch batch-1 decode engine for efficient token processing. - CUDA graph mode and megakernel (`--mega`) versions available for flexibility. Tech stack: Cuda Why it matters: The engine provides significant performance improvement on H100 GPUs, making it a valuable tool for developers working with Bonsai models. Repository: RightNow-AI/bonsai-turbo GitHub stars: 18

Overview: VSearcher is a multimodal search agent developed using reinforcement learning, designed to interact with real-world web environments and leverage up-to-date web information. It addresses the limitations of static multimodal models by enabling long-horizon, multi-turn tool use through iterative injection data synthesis and reinforcement learning. Key features: - Iterative Injection Data Synthesis pipeline to generate complex multimodal QA questions. - SFT-then-RL training pipeline to turn base multimodal models into multi-turn tool-callers. - Multimodal search benchmark MM-SearchExam for evaluating search capabilities of multimodal search agents. - Superior performance compared to recent multimodal search agents and proprietary models on multimodal web search tasks. Tech stack: Python, Reinforcement Learning, SFT-then-RL training pipeline, Iterative Injection Data Synthesis pipeline, Multi-turn tool calling, Multi-modal QA generation, Multi-modal web browsing, Multi-modal search capabilities evaluation. Why it matters: VSearcher demonstrates the potential of combining large models with reinforcement learning to create dynamic, interactive agents capable of handling complex tasks in real-world environments, thereby expanding the application scenarios of large models beyond text-only capabilities.

### Overview: This repository contains the code for reproducing the experiments on large-scale pre-training and transfer learning for the paper "Effect of large-scale pre-training on full and few-shot transfer learning for natural and medical images" (https://arxiv.org/abs/2106.00116). ### Key features: - Pre-trained models - Fine-tuning instructions - New datasets ### Tech stack: Jupyter Notebook ### Why it matters: This repository provides a comprehensive set of tools and instructions for researchers and developers interested in large-scale pre-training and transfer learning, specifically for applications in natural and medical imaging.

### Overview: 为解决电商客服在物流、售后等高频咨询场景下的痛点，搭建了Agent智能客服系统。通过意图分流、业务工具调用与RAG知识库检索，系统实现了订单物流查询、售后信息收集和商品问答等功能，显著提升了客服响应效率和回答准确性。 ### Key Features: - **订单物流查询**：通过意图识别与工具调用，快速获取订单状态和物流信息。 - **售后信息收集**：自动收集售后信息，确保问题处理的及时性和准确性。 - **商品问答**：使用RAG知识库检索，提供商品知识和手册问答，增强用户理解。 - **工具调用成功率**：99%的工具调用成功率，确保业务流程的高效执行。 - **意图识别准确率**：99%的意图识别准确率，有效分流用户问题。 - **知识检索召回率**：95%的检索召回率，提高知识匹配的准确性。 - **回答准确率**：93%的回答准确率，提升客服服务质量和用户体验。 ### Tech Stack: - Python - FastAPI - React - TypeScript - PostgreSQL - pgvector - RAG - Function Calling - Docker Compose ### Why it Matters: 通过智能客服系统，电商客服可以实现高效、准确的客户服务，减少人工重复查询和知识定位慢的问题，显著提升客户满意度和业务效率。

Overview: ScaleDP is an Open-Source library designed for processing documents using AI/ML capabilities and leveraging Apache Spark for scalability. Key features: - ✅ Processing PDF documents/Images to DataFrame - ✅ Extraction text/images from documents/Images - ✅ Zero-Shot structured data extraction using LLM and ML models - ✅ REST API service for minimal latency - ✅ Streaming mode for real-time processing Tech stack: Python Why it matters: ScaleDP simplifies document processing by integrating AI/ML capabilities with Apache Spark, making it efficient and scalable for various applications.

Overview: lorakit is a fast, config-driven toolkit for training Stable Diffusion XL (SDXL) using LoRA. It simplifies the process by requiring only a single YAML file. Key features: - One command, one config file - Fast: latent caching, optional `torch.compile`, and bf16 training - Fits your GPU: trains on 24 GB in bf16, or 4-bit QLoRA for 16 GB cards - Live previews during training - Resumable training - Flexible optimizers, LR schedulers, and LoRA targets for both UNet and text encoders - Optional latent face background preservation - Supports 19-class face parser Tech stack: Python Why it matters: lorakit streamlines the SDXL training process, making it accessible and efficient for developers and artists alike. Repository: TensorHarmony/lorakit GitHub stars: 19 GitHub description: Fast, config-driven SDXL LoRA & DreamBooth fine-tuning — train from a single YAML file.

Overview: This project is an art aesthetic style library that provides visual style decomposition into seven independent layers, allowing for easy swapping and integration with other styles, styles, and recipes. The library includes 421 styles, 274 hand-drawn styles, 100 director/film styles, and 157 camera-move recipes, each organized into seven distinct layers. Key features: - Visual style decomposition into seven independent layers - Integration with other styles, styles, and recipes - 421 styles, 274 hand-drawn styles, 100 director/film styles, and 157 camera-move recipes - 746 images (167 MB) of public domain images and hand-drawn style references - 15 guides, including a flowchart of 4 concepts, a keyword map, and a methodological approach Tech stack: HTML Why it matters: This project is designed to help developers and artists easily integrate and swap visual styles, making it a valuable tool for anyone looking to create visually cohesive projects or styles. The ability to easily swap and integrate styles allows for a more flexible and adaptable approach to style creation, making it a powerful addition to any developer's toolkit. Repository: YanKaFei/art-aesthetic-vault GitHub stars: 19 GitHub description: This project is an art aesthetic style library that provides visual style decomposition into seven independent layers, allowing for easy swapping and integration with other styles, styles, and recipes. The library includes 421 styles,

### Overview: This repository is a curated list of papers on **graph transfer learning (GTL)**, focusing on papers published in the years 2020 through 2023. It includes a variety of techniques and methods for improving graph neural networks in various domains. ### Key Features: - **Unified Pretraining for Recommendation via Task Hypergraphs** (WSDM 2024) - **A Data-centric Framework to Endow Graph Neural Networks with Out-of-Distribution Detection Ability** (KDD 2023) - **Graph Domain Adaptation via Theory-Grounded Spectral Regularization** (ICLR 2023) - **Multi-task Self-supervised Graph Neural Networks Enable Stronger Task Generalization** (ICLR 2023) - **Graph-Relational Domain Adaptation** (ICLR 2022) - **Cross-Domain Graph Anomaly Detection via Anomaly-aware Contrastive Alignment** (AAAI 2023) - **Graph Representation Learning for Multi-Task Settings: a Meta-Learning Approach** (IJCNN 2022) - **Graph Transfer Learning via Adversarial Domain Adaptation with Graph Convolution** (TKDE 2022) - **Cross-Domain Graph Anomaly Detection** (TNNLS 2022) - **Adaptive Transfer Learning on Graph Neural Networks** (KDD 2021) - **Transfer Learning of Graph Neural Networks with Ego-graph Information Maximization** (NIPS 2

Overview: TSDS is an implementation of the framework for selecting domain-specific and task-specific training data using an optimal-transport approach. Key features: - Optimized data selection for LLM finetuning and instruction tuning. - Supports both toy data and customized data. - Uses `faiss-gpu` for GPU acceleration if available. Tech stack: Python. Why it matters: TSDS enhances the efficiency and effectiveness of language model training by focusing on domain-specific and task-specific data, leading to better model performance and faster convergence.

### Overview: FlashSVD is an official implementation of "FlashSVD: Memory-Efficient Inference with Streaming for Low-Rank Models". This repository serves as the primary product path for FlashSVD v1.5, featuring lightweight and efficient streaming inference runtime for SVD-compressed language models. ### Key Features: - **Lightweight and Efficient**: Designed for low-rank transformers, FlashSVD is optimized for memory usage and inference speed. - **Streamable Inference**: Supports streaming inference, enabling real-time processing of large models. - **Cross-Model Integration**: Seamlessly integrates with Hugging Face's model integration, making it easy to use with existing models. - **CUDA/PyTorch Compatibility**: Supports both CPU and GPU environments, with a recommended environment bootstrap for GPU acceleration. ### Tech Stack: Python 3.10+ and a CUDA/PyTorch environment for GPU acceleration. ### Why It Matters: FlashSVD v1.5 offers significant performance improvements over traditional SVD-based models, achieving up to 1.53x end-to-end decode speedup on the Llama-2-7B configuration. This makes it a valuable tool for developers and researchers working with low-rank transformers, accelerating their workflows and improving model performance.

### Overview: Typhoonminigen is a macOS image generator for FLUX.2 Klein, built entirely on Apple Silicon with MLX, offering a native, on-device experience without Python, ComfyUI, or local servers. It supports various language prompts and includes an on-device LoRA training studio, a 247-scene library, and a gallery with reproducible recipes. ### Key features: - **On-Device Generation**: Fully native on Apple Silicon, eliminating the need for a local server or Python. - **MLX Integration**: Utilizes MLX for efficient model inference and training. - **Swift Development**: Built with Swift 6.0+ and supports macOS 14+ and Apple Silicon. - **No Python Dependencies**: No need for Python to run the app. - **LoRA Training Studio**: A dedicated studio for on-device LoRA training. - **247-Scene Library**: A library of 247 scenes for generating images. - **Image-to-Image and LoRA Adapter Support**: Supports image-to-image generation and LoRA adapters. - **Real-Time Telemetry**: Provides real-time system telemetry for better performance and stability. ### Tech stack: Typhoonminigen is built with Swift and MLX, leveraging Apple Silicon for efficient model inference and training. ### Why it matters: Typhoonminigen offers a native, on-device experience for FLUX.2 Klein, providing a seamless and efficient workflow without the need for external dependencies or local servers. It supports various language prompts and includes advanced features like an on-device LoRA training

### Overview: This project focuses on creating a multi-modal dataset for hate speech detection, combining text and images. The goal is to prepare a clean, structured dataset that can be used to train a deep learning model capable of detecting hate speech in multi-modal content. ### Key Features: - Multi-modal dataset preparation - Text cleaning and preprocessing - Image preprocessing - Multi-modal dataset creation - DataLoader preparation - Multi-modal embeddings generation - Model training on generated embeddings ### Tech Stack: Jupyter Notebook ### Why it Matters: This project is crucial for developing a robust deep learning model that can effectively identify hate speech across various media platforms, ensuring a more comprehensive and accurate detection system.