Projects

Overview: GEO（Generative Engine Optimization）智能内容优化平台是一款帮助品牌在AI搜索引擎中被优先、准确、可信地提及的工具，旨在提升品牌曝光、优化搜索排名、建立权威形象并提供数据驱动的内容策略。 Key features: - 关键词策略：智能关键词生成、搜索意图覆盖、数据增强、语义扩展 - 内容生成：20+平台模板、E-E-A-T内嵌、品牌植入策略、结构化输出 - 知识库（RAG）：文档管理、智能检索、来源保障 - 内容优化：E-E-A-T强化、事实密度增强、结构化数据、内容评分 - 多模型验证：DeepSeek、OpenAI、通义千问等多模型交叉验证 - AI搜索验证：接入Perplexity真实搜索引擎验证 - 竞品对比：同时验证自有品牌和竞品 - 负面防护：主动生成负面查询并监控 - 数据分析：提及率趋势、平台贡献度、ROI分析、内容独特性 - 平台同步：一键发布到GitHub、12平台 - 工作流自动化：一键完成全流程 - 资源库：优化资源推荐、品牌知识库 Tech stack: Python, Streamlit, LangChain, GitHub API Why it matters: 通过GEO平台，品牌能够显著提高在AI搜索引擎中的曝光率，优化搜索排名，建立权威形象，并通过数据驱动的内容策略提升品牌表现。

Overview: This repository is an awesome list of vector databases, covering both open-source and commercial solutions for storing, managing, and searching high-dimensional data. It includes Faiss, Annoy, Milvus, HNSWLIB, NMSLIB, Cottontail DB, Vexvault, Pinecone, Vespa, Vector.ai, Qdrant, Weaviate, and NucliaDB. Key features: - Faiss: A library for efficient similarity search and clustering of dense vectors, supporting GPU and CPU, indexing methods like LSH, PQ, HNSW, and more. - Annoy: An approximate nearest neighbor search library for fast query times, supporting static and dynamic datasets, and custom distance metrics. - Milvus: An open-source vector database for AI and analytics, supporting multiple index types, hybrid search, and RESTful API and SDKs. - HNSWLIB: An open-source C++ library for approximate nearest neighbor search, implementing the Hierarchical Navigable Small World algorithm. Tech stack: C++, Python, JavaScript, Java, C#, Docker, and more. Why it matters: These vector databases are crucial for storing, managing, and efficiently searching large volumes of high-dimensional data in applications like machine learning, artificial intelligence, and data-intensive analytics. Repository: mileszim/awesome-vector-database GitHub stars: 182 GitHub description: A curated list of awesome vector databases for storing, managing, and searching high-dimensional data.

Overview: Coltt is a vector database designed for efficient and scalable search, supporting Multi-Vector Search, HNSW, FLAT, and quantization. It enables fast searches through sophisticated shard design, making it suitable for edge environments and small-scale production settings. Key features: - Multi-Vector Search - High-performance HNSW and FLAT - Advanced internal data shard design Tech stack: Go Why it matters: Coltt is optimized for real-time search and retrieval, making it ideal for applications requiring fast and accurate results in various environments, from edge nodes to small-scale production setups. Repository: sjy-dv/coltt GitHub stars: 182 GitHub description: Coltt is a database designed to be implemented from scratch to production. It supports Multi-Vector Search, high-performance HNSW, FLAT, and quantization, and enables fast searches through sophisticated internal data shard design. It is envisioned and developed to be used reliably in large-scale production environments as well. Primary language: Go GitHub description: Coltt is a database designed to be implemented from scratch to production. It supports Multi-Vector Search, high-performance HNSW, FLAT, and quantization, and enables fast searches through sophisticated internal data shard design. It is envisioned and developed to be used reliably in large-scale production environments as well.

### Overview: This project implements a practical membership inference attack against fine-tuned large language models using self-prompt calibration. The proposed method, SPV-MIA, is designed to be applied to various large language models and fine-tuned on different datasets. ### Key features: - **Self-prompt Calibration (SPV-MIA)**: A membership inference attack that leverages self-prompted fine-tuning to infer the membership of specific tokens within a large language model. - **Fine-tuning of Target Model**: The project includes fine-tuning of large language models using the transformers library, which can be fine-tuned on any dataset using the accelerate library. - **Self-prompt Reference Model Fine-tuning**: Before fine-tuning the self-prompt reference model, a reference dataset can be sampled using a self-prompt approach. - **Dependency Installation**: The project requires specific dependencies such as torch, accelerate, transformers, and others, which can be installed using pip. ### Tech stack: - Python ### Why it matters: This project is significant for researchers and developers interested in membership inference attacks on large language models, as it provides a practical and effective method to test the robustness of fine-tuned models against membership inference attacks.

Overview: This repository contains community plugins for Google Firebase Genkit, built by the community. It includes plugins for various AI and vector store APIs, such as Anthropic, Cohere, Groq, Mistral, OpenAI, Azure OpenAI, Convex, HNSW, Milvus, and more. Key features: - Community-built plugins for Google Firebase Genkit - Support for AI models and vector stores - Examples available in the examples directory Tech stack: TypeScript Why it matters: This repository provides developers with community-contributed plugins for integrating AI and vector store functionalities into their Firebase Genkit projects, making it easier to implement advanced features without needing to build everything from scratch. Repository: BloomLabsInc/genkit-plugins Primary language: TypeScript GitHub stars: 181

Overview: ScienceAgentBench is a benchmark designed to rigorously assess language agents for data-driven scientific discovery. It includes 102 tasks extracted from peer-reviewed papers across four disciplines, validated by subject matter experts, and evaluated using a comprehensive set of metrics. Key features: - 102 tasks extracted from 44 peer-reviewed publications - 9 subject matter experts for validation - Comprehensive evaluation metrics for generated programs, execution results, and costs - Multiple rounds of manual validation by annotators and subject matter experts Tech stack: Python Why it matters: ScienceAgentBench provides a reliable and scientific benchmark for evaluating language agents in data-driven scientific discovery, ensuring that the outcomes are both authentic and relevant to real-world applications. Repository: OSU-NLP-Group/ScienceAgentBench Primary language: Python GitHub stars: 179 GitHub description: [ICLR'25] ScienceAgentBench: Toward Rigorous Assessment of Language Agents for Data-Driven Scientific Discovery GitHub description excerpt: Code and data for the paper "[ScienceAgentBench: Toward Rigorous Assessment of Language Agents for Data-Driven Scientific Discovery](https://arxiv.org/abs/2410.05080)". Repository updates: - 2026/04/30: Verified version released to mitigate false negatives in evaluation - 2025/01/10: Containerized evaluation harness released - 2024/12/10: Curated

Overview: BlahST is a lean, fast, and accurate speech-to-text input tool for Linux, allowing you to type text from your microphone directly into any editable text field. It utilizes whisper.cpp offline and llama.cpp for conversational AI interactions. Key features: - Speech recognition with whisper.cpp - Language selection and translation with wsiml - Downloadable whisperfile for offline use - Continuous speech input with blooper - Interaction with local LLMs via llama.cpp or llamafile - AI proofreading and multilingual speech chat Tech stack: Shell Why it matters: BlahST offers a fast, efficient, and powerful way to input text from speech, making it a valuable tool for developers and anyone needing to quickly type text into Linux environments. Repository: QuantiusBenignus/BlahST GitHub stars: 181 GitHub description: Input text from speech in any Linux window, the lean, fast and accurate way, using whisper.cpp OFFLINE. Speak with local LLMs via llama.cpp.

### Overview: Context Engineering: Build Consistent, Accurate, Predictable AI Systems is a comprehensive guide to understanding and implementing context engineering principles, focusing on designing systems that provide a comprehensive and structured context for large language models (LLMs) and AI agents to complete tasks effectively. ### Key features: - **Context Engineering Principles**: A foundational understanding of context engineering, including instructions, external knowledge, tools, memory, and state. - **Hands-On Examples**: Practical examples in various programming languages (Python, JavaScript, Java) to illustrate context engineering concepts. - **Comprehensive Table of Contents**: A detailed table of contents for the book, covering introduction, instructions, external knowledge, tools, memory, user prompts, context management, evaluation, governance, and AI frameworks. - **GitHub Repository**: A companion resource for the book, providing examples and references for context engineering principles. ### Tech stack: Python ### Why it matters: The success of complex LLM-based systems depends more on the quality and completeness of the context provided than on the specific wording of the prompt itself. This book aims to provide a strong theoretical foundation and practical examples to help practitioners understand and adopt context engineering principles.

### Overview: Shot2Story is a comprehensive video understanding benchmark that includes detailed long summaries and shot-level captions for multi-shot videos. This project aims to enhance the ability of AI systems to comprehend and summarize videos effectively. ### Key features: - Detailed video text summaries - Shot-level captions - Comprehensive dataset of 134K videos and 188K video shots - Updated 134K version with 43K human annotations and 90K GPTV-generated summaries - Video textual summary generation demo - Online chatbot demo for video summarization ### Tech stack: Python ### Why it matters: Shot2Story provides a valuable benchmark for evaluating video summarization models, offering insights into temporal-related, holistic-understanding, and audio-related aspects of video comprehension. The detailed summaries and shot-level captions enhance the accuracy and relevance of video summaries, making it a crucial resource for researchers and developers in the field of video understanding and AI.

### Overview: v100-skinny is a software implementation of Qwen3.8-27B's mixed NVFP4/FP8 weight representation, running on four Tesla V100-SXM2-16GB GPUs. It leverages Volta's `mma.sync.m8n8k4` for efficient execution, achieving a 35% longer round with 38% more tokens committed per round, making it suitable for the model's native MTP depth. ### Key features: - Hand-written NVFP4 W4A16 CUDA kernels for Volta. - Runs QPN2 (W4A16) and QPN8 (W8A16) on Volta’s `mma.sync.m8n8k4`. - No FP8→FP4 requantization or persistent FP16 weight copy. - Supports FP16 activations and KV cache. - OpenAI-compatible serving. ### Tech stack: Python ### Why it matters: This implementation allows for a more efficient and accurate inference on the Qwen3.8 model, leveraging Volta's capabilities without the need for hardware acceleration, thus reducing costs and power consumption.

### Overview: This repository contains summaries of various papers on deep learning, focusing on historical context, fundamental concepts, and recent advancements. The summaries are organized by week and year, providing a comprehensive overview of the field's development and applications. ### Key Features: - **Historical Context**: Summaries of key papers from the inception of deep learning to present-day advancements. - **Fundamental Concepts**: Detailed explanations of key concepts such as deep neural networks, convolutional neural networks, and recurrent neural networks. - **Applications**: Summaries of deep learning applications in natural language processing, computer vision, robotics, and more. - **Techniques**: Coverage of advanced techniques like neural Turing machines, reinforcement learning, and generative models. ### Tech Stack: - **Python**: Used for summarizing and organizing the papers. - **Markdown**: Used for formatting the README and creating the summaries. ### Why It Matters: Deep learning has revolutionized numerous fields, from computer vision to natural language processing. This repository serves as a valuable resource for understanding the evolution and current state of deep learning, providing insights into the techniques and applications that have shaped the field. ### Repository: - **Repository**: gopala-kr/summary - **Stars**: 178 - **Description**: Summaries of all the papers I read on deep learning. ### Articles: - **Jan 2018 - March 2018**: Summaries of articles on deep learning and related topics. - **April 2018 - June 2018**: Summaries of articles on deep

### Overview: AutoTools is an official project for the WWW 2025 paper "Tool Learning in the Wild: Empowering Language Models as Automatic Tool Agents". It aims to introduce an automatic framework that enables large language models (LLMs) to act as autonomous agents by transforming tool documentation into executable functions. This project is designed to solve practical tasks by integrating the AutoTools with the LLMs. ### Key Features: - Tool Documentation to Function Transformation - Syntax and Runtime Verification - Flexible Function Composition ### Tech Stack: Python ### Why It Matters: AutoTools provides a practical solution for integrating language models with practical tools, making them more versatile and efficient in their applications. This framework enhances the capabilities of language models by enabling them to perform tasks that require specific tool functionalities, thereby expanding their utility in various domains. ### Repository: mangopy/AutoTools ### GitHub Stars: 181 ### GitHub Description: Official Repo for WWW 2025 paper "Tool Learning in the Wild: Empowering Language Models as Automatic Tool Agents" ### Quick Start: AutoTools is designed to be easily integrated into existing LLMs. Here’s how you can get started: 1. **Build Your Own Tool-Augmented (Func-Augmented) LLMs:** ```python from basellm import FuncLLM llm_config = { "model_name_or_path": "<MODEL_NAME>", "base_model": "gpt", # gpt or open-source llm "api_key": "<API-KEY>", #

### Overview: This project aims to create three types of chatbots for mental health using Python, with the goal of making mental health more accessible and improving patient-clinician interactions. The chatbots are designed based on different attributes and are trained on a dataset of FAQs about mental health. ### Key Features: - **Rule-Based Chatbot**: Uses TF-IDF and NLTK tokenizer for data preprocessing and cosine similarity for evaluation. - **Retrieval-Based Chatbot**: Trained on Machine Learning and Deep Learning models like RNN, LSTM, Bi-LSTM, GRU, and CNN. - **Generative-Based Chatbot**: Utilizes Natural Language Processing (NLP) to mimic human conversation patterns. ### Tech Stack: Python ### Why It Matters: The project aims to develop chatbots that can be used in mental health interventions, providing a more accessible and effective way to address mental health concerns.

Overview: This project provides a detailed breakdown of knowledge graph RAG, including diagrams and examples, focusing on how it addresses the limitations of traditional RAG systems by introducing a structured, hierarchical approach to information organization and retrieval. Key features: - Diagrams and examples of knowledge graph RAG - Explanation of how knowledge graphs capture natural associations between concepts - Introduction of LLMs and their capabilities in constructing knowledge graphs - Comparison with traditional vector similarity techniques Tech stack: Jupyter Notebook Why it matters: This project offers a practical demonstration of how knowledge graph RAG can be used to enhance semantic understanding and reasoning in natural language processing, making it a valuable tool for developers looking to improve their AI applications.

### Overview: Nightmare Mobile is an on-device, ComfyUI-style node graph for image and video diffusion on Android. It runs on the Qualcomm Hexagon NPU, enabling text-to-image, image-to-image, inpainting, upscaling, and more directly on the phone. The app supports SD1.5, SDXL, Flux, Z-Image, Anima, Qwen Image 2.1, Krea 2, and MOSS Image, among others. ### Key features: - **Built for a Phone**: The app is designed specifically for Android devices, featuring big ports, pinch-to-zoom, and a run bar that folds down for a larger canvas. - **Local Dream Models**: Local Dream's models are kept in the `Download/LocalDream` folder and can be run directly on the phone without a second download. - **SDXL without NPU**: Local Dream's SDXL models can be rendered on the CPU/GPU, making it suitable for devices with an NPU that cannot run SDXL. - **UltraFix**: A tiled repair feature that redraws the detail of an upscaled picture without changing it. - **Scheduler, Seed, and CFG**: These settings can be toggled on the Image node, but they require a download and are not available in the folded options. ### Tech stack: Kotlin ### Why it matters: Nightmare Mobile offers a powerful and efficient image and video diffusion solution directly on an Android device, without the need for a server or account. It's ideal for users who prefer a fully offline experience and want to leverage

### Overview: OpenPronounce is a self-hosted alternative to Azure Pronunciation Assessment that provides accurate phoneme-level pronunciation assessment using Wav2Vec2 and DTW. It offers a user-friendly command-line interface and a web app, allowing developers to evaluate their pronunciation skills without the need for an API key or monthly fees. ### Key features: - **Command Line Interface (CLI):** Record a pronunciation and receive detailed feedback on pronunciation accuracy, including scores, transcriptions, and mispronounced words with their correct phonemes. - **Web App:** A web-based interface for easy access, allowing users to record their pronunciation, get scores, and see mispronounced words highlighted. - **Self-Hosted:** Runs entirely on the user's machine, eliminating the need for external APIs or billing. - **Language Support:** Supports English, French, Spanish, German, Italian, Portuguese, and Dutch. ### Tech stack: The project uses Python, Wav2Vec2, DTW, and other open-source tools. It is self-hosted and runs on the user's machine, making it ideal for language learning apps, EdTech products, and research. ### Why it matters: OpenPronounce is a valuable tool for language learners and educators, offering a reliable and cost-effective way to assess pronunciation skills without the need for external APIs or expensive services.

### Overview: M-Courtyard is a desktop assistant designed for fine-tuning local AI models on Apple Silicon. It simplifies the process of fine-tuning and data preparation, ensuring privacy and zero-code execution. ### Key Features: - **Zero-Code Pipeline**: Fine-tuning, data preparation, and export in 4 easy steps. - **100% Local & Private**: No data leaves your machine, perfect for sensitive data. - **Optimized for MLX**: Uses `mlx-lm` to maximize performance on M1/M2/M3/M4 chips. - **AI-Powered Data Prep**: Automatically converts documents into datasets using local models or fallback to built-in rules. ### Tech Stack: TypeScript ### Why It Matters: M-Courtyard simplifies the process of fine-tuning and data preparation for local AI models, ensuring privacy and ease of use on Apple Silicon.

### Overview: This repository contains the code for the COLM 2025 paper "Critique Fine-Tuning: Learning to Critique is More Effective than Learning to Imitate." It introduces Critique Fine-Tuning (CFT), a paradigm shift in language model training where models learn to critique rather than imitate. ### Key features: - Implement Critique Fine-Tuning (CFT) for language model training. - Provide a dataset and models for training and evaluation. - Include a self-construct critique data generation script. ### Tech stack: Python ### Why it matters: This code represents a significant advancement in language model training, offering a more effective approach to fine-tuning models for specific tasks.

### Overview: ImagenHub is a comprehensive library designed to standardize the evaluation and inference of conditional image generation models. It includes 7 prominent tasks with 7 high-quality datasets, a unified inference pipeline, and supports over 50 models. ImagenHub also features two human evaluation scores, detailed guidelines, and provides tools for visualization, autometrics, and Amazon Mechanical Turk templates. ### Key features: - **7 Prominent Tasks**: ImagenHub covers 7 key tasks in conditional image generation. - **7 High-Quality Datasets**: Each task is accompanied by curated datasets for fair comparison. - **Unified Inference Pipeline**: Ensures fair and consistent evaluation across different models. - **50+ Models Supported**: A robust ecosystem for various conditional image generation models. - **Two Human Evaluation Scores**: Semantic Consistency and Perceptual Quality scores with comprehensive guidelines. - **Visualization, Autometrics, and MTurk Templates**: Tools for easy visualization, automated metrics, and Amazon Mechanical Turk integration. ### Tech stack: Python ### Why it matters: ImagenHub simplifies the evaluation and inference process for conditional image generation models, making it easier for researchers and practitioners to compare and evaluate different models. Its comprehensive features and support for a wide range of models make it a valuable tool in the field of visual AI research and development.

### Overview: This repository is the official implementation of the paper "Move as You Say, Interact as You Can: Language-guided Human Motion Generation with Scene Affordance". It introduces a novel two-stage framework that leverages scene affordance as an intermediate representation, effectively linking 3D scene grounding and conditional motion generation. The method, based on the Affordance Diffusion Model (ADM) and Affordance-to-Motion Diffusion Model (AMDM), addresses the challenges of generating language-guided human motion in 3D environments. The repository includes preprocessed data, extensive experiments, and a detailed project page. ### Key Features: - **Affordance Diffusion Model (ADM)**: Predicts explicit affordance maps. - **Affordance-to-Motion Diffusion Model (AMDM)**: Generates plausible human motions. - **Scene Affordance Representation**: Links 3D scene grounding and conditional motion generation. ### Tech Stack: - Python ### Why It Matters: This method significantly advances the field of text-to-motion synthesis by providing a powerful framework that can generate language-guided human motion within 3D environments, overcoming the limitations of existing models. The use of scene affordance as an intermediate representation allows for more effective and generalizable motion generation, making it a valuable contribution to the field of human motion generation.