Projects

### Overview: TopicGPT is a Python framework designed for topic modeling, specifically focusing on the NAACL'24 paper "TopicGPT: Topic Modeling by Prompting Large Language Models." This project includes a package `topicgpt_python` that automates the process of generating, refining, and assigning topics to input text, utilizing various APIs for different deployment options. ### Key Features: - **Generate High-Level Topics**: `generate_topic_lvl1` function creates high-level, generalizable topics. - **Generate Low-Level Topics**: `generate_topic_lvl2` function generates low-level, specific topics for each high-level topic. - **Refine Topics**: `refine_topics` function merges similar topics and removes irrelevant ones. - **Assign Topics**: `assign_topics` function assigns generated topics to input text, along with a supporting quote. - **Correct Topics**: `correct_topics` function corrects the generated topics by reprompting the model to ensure the final topic assignment is grounded in the topic list. ### Tech Stack: - Python 3.9+ - `topicgpt_python` package - OpenAI API, Vertex AI, Azure API, Gemini API, vLLM (requires GPUs for inference) ### Why It Matters: TopicGPT simplifies the process of topic modeling by leveraging the power of large language models, making it accessible and efficient for researchers and practitioners alike. The framework ensures that the generated topics are both high-level and specific, providing a comprehensive understanding of the input text.

### Overview: This repository contains a minimal implementation of diffusion models for text generation. It learns a diffusion model from a given text corpus, allowing for the generation of text samples from the learned model. ### Key features: - Diffusion models for text generation - Simple corpus for quick iterations and testing - Tokenizer for text processing - Training and inference scripts available ### Tech stack: Python ### Why it matters: This implementation provides a straightforward way to experiment with diffusion models for text generation, making it accessible for developers and researchers.

### Overview: 文枢是一个深度上下文感知的智能体小说创作系统，旨在通过拆分长篇写作过程为多个可见、可维护、可回溯的模块，提供清晰稳定的写作工作台。前端提供清晰稳定的写作工作台，后端通过orchestrator、agents、context_engine、storage等组件组织创作流程，项目数据存储在YAML、Markdown、JSONL等纯文本格式中，适合版本管理和长期维护。 ### Key features: 1. **清晰的分卷章节结构**：文枢通过显式维护“卷 -> 章节 -> 草稿/摘要”的写作结构，确保写作过程中的顺序和边界清晰。 2. **卡片系统构建人物/世界观/文风设定集**：将长期设定资产拆成三类核心卡片，提供可维护、可检索、可复用的项目资产。 3. **按距离索引的事实摘要系统**：通过事实主存储在`canon/facts.jsonl`，章节摘要与分卷摘要分别保存为YAML，结合BM25、实体增强、章节绑定与事实距离衰减来挑选相关上下文。 4. **同人创作功能**：通过搜索、预览、抓取、提案等步骤，支持同人创作，提供提案形式的原作世界观。 5. **支持主流模型供应商**：模型层通过统一抽象，支持OpenAI、Anthropic、DeepSeek、Gemini、Qwen通义千问、Wenxin文心一言

Overview: Open RAG Eval is an open-source Python toolkit designed to evaluate Retrieval-Augmented Generation (RAG) systems without the need for "golden answers." It provides a flexible framework for measuring RAG quality and includes metrics from the TREC-RAG benchmark, connectors for Vectara and LlamaIndex, and options for optional golden answer evaluation. Key features: - Standard Metrics: Provides TREC-RAG evaluation metrics. - Modular Architecture: Allows easy integration of custom metrics and connectors. - Detailed Reporting: Generates per-query scores and intermediate outputs. - Visualization: Supports comparison across different configurations or runs. Tech stack: Python Why it matters: RAG evaluation can be complex, and Open RAG Eval simplifies the process by providing a flexible and extensible framework, making it easier and more scalable to evaluate RAG systems. Repository: vectara/open-rag-eval GitHub stars: 418

### Overview: This repository is a comprehensive list of resources for learning the mathematics required for machine learning and artificial intelligence. It covers essential topics such as algebra, linear algebra, calculus, and statistics, providing a structured pathway for beginners and those seeking to refresh their knowledge. ### Key Features: - **Algebra**: Includes resources for learning algebra from Khan Academy. - **Linear Algebra**: Covers books and MOOCs on linear algebra, including Gilbert Strang's MIT OCW course. - **Calculus**: Features MOOCs and books on calculus, focusing on specific applications relevant to machine learning. - **Statistics & Probability**: Provides MOOCs and books on statistics and probability, including "Think Stats" by Allen Downey. ### Tech Stack: - **GitHub**: 418 stars, indicating a large community engagement and active usage. ### Why It Matters: Mathematics is crucial in machine learning as it forms the foundational principles that underpin the algorithms and models used in AI systems. Understanding these mathematical concepts is essential for anyone looking to build or work with machine learning models effectively.

Overview: This project is a contextual word checker designed to provide better suggestions for spelling errors, focusing on non-word errors (NWE) using a BERT model. It aims to improve the package by extending functionality to identify real word errors (RWE) and optimize the package. Key features: - Contextual Spell Check: Provides suggestions for spelling errors based on context. - Non-Word Error (NWE) Correction: Focuses on correcting non-word errors using a BERT model. - Real Word Error (RWE) Identification: Plans to extend functionality to identify real word errors. - Python 3.6+ Compatibility: The package is compatible with Python 3.6 and above. Tech stack: The project is built using Python, with the primary language being Python. Why it matters: This package is essential for improving the accuracy of spelling suggestions, especially for non-word errors, by leveraging the context provided by a BERT model. It aims to enhance the quality of suggestions and potentially improve the overall user experience by identifying and correcting real word errors more effectively. Repository: The project is hosted on GitHub at [R1j1t/contextualSpellCheck](https://github.com/R1j1t/contextualSpellCheck). GitHub stars: The project has received 418 GitHub stars, indicating a high level of interest and engagement from the developer community. Primary language: The project is written in Python. GitHub description: The project is a contextual word checker for better suggestions, focusing on non-word errors using a BERT model. It aims to improve the package by extending functionality

### Overview: Artificio is a comprehensive suite of deep learning computer vision algorithms designed for practical applications in real-world scenarios. ### Key features: - **Image Retrieval (via Transfer Learning)**: Enhances image retrieval capabilities using pre-trained models. - **Image Retrieval (via Autoencoders)**: Utilizes autoencoders for efficient image retrieval. - **Image Processing Tools**: Offers a range of tools for image processing and manipulation. ### Tech stack: Python ### Why it matters: Artificio provides a robust set of deep learning algorithms tailored for real-world applications, ensuring efficient and effective computer vision solutions.

Overview: CLLMs, or Consistency Large Language Models, are a family of models that reduce inference latency by efficiently decoding multiple tokens in parallel. This method, called Jacobi decoding, improves inference efficiency compared to conventional auto-regressive decoding. Key features: - Jacobi decoding for efficient parallel decoding - Reduced inference latency - Achieved $2.4\times$ to $3.4\times$ speed improvements on various tasks Tech stack: Python Why it matters: CLLMs offer a significant speedup in inference time, making them ideal for applications requiring fast generation, such as chatbots, language models, and more. Repository: hao-ai-lab/Consistency_LLM Primary language: Python GitHub stars: 418 GitHub description: [ICML 2024] CLLMs: Consistency Large Language Models README excerpt: [Image and Text Snippet] Contents: - News - Introduction - Installation - Model Weights - Usage (Inference, Training, Evaluation) - Citation News: - CLLMs integrated in FastChat - CLLM paper available on arXiv Introduction: CLLMs are a family of efficient parallel decoders derived from pre-trained language models. Compared to existing fast decoding techniques, CLLMs eliminate the need for draft models and architectural modifications, offering seamless integration with other techniques for improved inference speed. Installation: 1. Set up environment: `conda create -n cllm python=3.10` 2. Clone repository: `git clone git@github

Overview: DeLFT is a Deep Learning Framework for Text designed to process rich text and sequence labeling tasks, offering a comprehensive solution for both sequence labeling and text classification. Key features: - Implements modern NLP architectures and tasks with a consistent API and input formats. - Reduces the size of RNN models by removing word embeddings. - Supports categorical features in various architectures. - Uses dynamic data generators for efficient training. - Efficiently manages pre-trained embeddings. - Provides a comprehensive evaluation framework with standard metrics for sequence labeling and classification tasks. Tech stack: Python Why it matters: DeLFT offers a practical and efficient solution for NLP tasks, focusing on sequence labeling and text classification, while also supporting rich text processing and reproducibility through benchmarking and production-level integration.

Overview: MCP Apple Notes is a Model Context Protocol (MCP) server designed for integrating Claude with your Apple Notes, enabling semantic search and RAG capabilities. Key features: - 🔍 Semantic search over Apple Notes using `all-MiniLM-L6-v2` on-device embeddings model - 📝 Full-text search capabilities - 📊 Vector storage using LanceDB - 🤖 MCP-compatible server for AI assistant integration - 🍎 Native Apple Notes integration via JXA - 🏃‍♂️ Fully local execution - no API keys needed Tech stack: TypeScript Why it matters: MCP Apple Notes simplifies the integration of Claude with Apple Notes, allowing for seamless semantic search and RAG-based interactions, making it easier to integrate AI assistants into existing workflows. Repository: RafalWilinski/mcp-apple-notes GitHub stars: 417

Overview: Belajarpython adalah situs tutorial pemrograman Python Indonesia yang dikembangkan oleh komunitas developer. Key features: - Mulai belajar bahasa pemrograman Python dari dasar - Baca artikel terbaru dan terpopuler tentang Python - Ajukan pertanyaan tentang Python Tech stack: Nunjucks Why it matters: Belajarpython membantu pemula mempelajari bahasa pemrograman Python dengan tutorial yang mudah dipahami dan artikel yang disusun oleh komunitas.

Overview: Dragonfly is a Python-based speech recognition framework that allows for powerful scripting and extension of speech recognition software like Dragon NaturallySpeaking, Windows Speech Recognition, Kaldi, and CMU Pocket Sphinx. Key features: - Customizable speech commands and grammar objects - Cross-platform compatibility (Windows, macOS, Linux (X11 only)) - Actions framework for text input and key-stroke simulation - Support for multiple speech recognition engines: Dragon, Windows Speech Recognition, Kaldi, and CMU Pocket Sphinx Tech stack: Python Why it matters: Dragonfly simplifies integration of speech recognition with Python applications, making it easier to automate tasks and create custom commands for speech-enabled software.

Overview: ExplainToMe is an automatic text summarizer that uses the TextRank algorithm to generate concise machine-generated summaries from web articles. Key features: - Automatically extracts key points from web articles - Supports multiple languages, including English, Korean, and Japanese - Integrates with Docker and Kitematic for easy deployment Tech stack: Python Why it matters: This tool simplifies the process of summarizing complex web articles, making it easier for developers to quickly understand key information from various sources.

### Overview: This repository contains a collection of papers and notes on various aspects of Natural Language Processing (NLP), focusing on distributed representations and sentence embeddings. ### Key Features: - **Distributed Word Representations**: Papers on improving vector space word representations using multilingual correlation, visualizing data using t-SNE, and compositional character models. - **Distributed Sentence Representations**: Research on enriching word vectors with subword information. - **Entity Recognition (Sequence Tagging)**: Papers on sequence tagging and language modeling. - **Language Model (LM for pre-training)**: Techniques for pre-training language models. - **Machine Translation**: Papers on machine translation models. - **Question Answering (Machine Reading Comprehension)**: Research on question answering and machine reading comprehension. - **Recommendation Systems**: Papers on recommendation systems. - **Relation Extraction**: Techniques for extracting relationships between entities. - **Sentences Matching (Natural Language Inference/Textual Entailment)**: Research on natural language inference and entailment. - **Text Classification (Sentiment Classification)**: Papers on sentiment classification. - **Materials/Toolkits/Corpus**: Resources and tools for NLP. ### Tech Stack: - Python, TensorFlow, PyTorch, and other machine learning libraries. ### Why it Matters: These papers and notes provide a comprehensive overview of the latest advancements in NLP, focusing on how to represent and understand language through distributed representations. They are essential for researchers and practitioners looking to improve their NLP models and algorithms.

### Overview: Anymove is a motion capture tool that converts one AI video into an animation for any rigged creature, including wings, tails, and eight legs. It leverages the expertise of [@CanItCode] and [@Graalitoo] to create a working version of their methods, including the MiniMax H3 LoRA step. ### Key Features: - **Motion Capture from AI Video**: Anymove captures motion from one AI video and converts it into a detailed animation. - **Rigged Creature Support**: Supports creatures with wings, tails, and eight legs. - **Two Cameras in One Frame**: Captures motion from two cameras in one frame, providing a comprehensive view of the creature. - **Joint Reader and GLB Export**: Uses a trained joint reader to find and pin joints in both camera views, and exports the rig as a GLB file. - **Minimalist Approach**: The method is based on the expertise of [@CanItCode] and [@Graalitoo], with the MiniMax H3 LoRA step. ### Tech Stack: Python ### Why It Matters: Anymove simplifies the process of creating detailed animations from AI videos, making it accessible for anyone interested in motion capture and rigging. The tool ensures accurate and detailed motion capture, making it valuable for various applications in animation and game development.

### Overview: This project is the codebase for the book "AI-Powered Search" by Trey Grainger, Doug Turnbull, and Max Irwin. It provides comprehensive examples and tutorials on modern machine learning techniques for building intelligent search engines. The book covers topics such as semantic search, retrieval augmented generation, question answering, and more, with a focus on integrating large language models like GPT. The project is designed to be run using Docker containers, making it easy to set up and run locally. ### Key Features: - Semantic search using dense vector embeddings from foundation models - Retrieval Augmented Generation - Question answering and summarization combining search and LLMs - Fine-tuning transformer-based LLMs - Personalized search based on user signals and vector embeddings - Collecting user behavioral signals and building signals boosting models - Semantic knowledge graphs for domain-specific learning - Implementing machine-learned ranking models (learning to rank) - Building click models to automate machine-learned ranking - Generative search, hybrid search, and the search frontier ### Tech Stack: Python with PySpark ### Why it Matters: This book provides a practical guide for developers and researchers to build search engines that are intelligent and domain-aware. By leveraging modern machine learning techniques, it prepares readers to build search engines that can understand natural language queries, user preferences, and context, delivering more relevant and user-friendly search experiences.

Overview: Neural Monkey is an open-source tool designed for sequence learning in Natural Language Processing (NLP), built on TensorFlow. It simplifies the process of creating and using neural network models for tasks like neural machine translation and sentence classification. Key features: - High-level abstraction for sequential neural network models - Built on TensorFlow - Supports fast prototyping of NLP models - Includes standard building blocks like RNN encoders and decoders, and multi-layer perceptrons - Provides a simple way to add new building blocks directly in TensorFlow Tech stack: Python Why it matters: Neural Monkey offers a user-friendly interface for building and deploying NLP models, making it accessible to developers and researchers alike. Its flexibility and ease of use make it an excellent choice for anyone looking to implement advanced NLP techniques quickly and efficiently.

### Overview: A Python project featuring YURI from the game Doki Doki Literature Club, where she accesses the webcam and stares directly into the player's soul. ### Key Features: - **Face Tracking**: Tracks the player's facial expressions and adjusts YURI's gaze accordingly. - **Proximity Detection**: YURI blushes when the player gets close to the webcam. - **Customizable Characters**: Users can replace YURI with their own anime characters by providing two images: a character without eyes and eyes. ### Tech Stack: Python 3.5 or above, OpenCV, NumPy, PyGame. ### Why It Matters: This project combines elements of anime and computer vision, offering a unique and engaging experience for users.

Overview: AgentLego is an open-source library designed to enhance Large Language Model (LLM) based agents with a rich set of tool APIs, including visual perception, image generation and editing, speech processing, and visual-language reasoning. It supports easy integration with popular agent frameworks like LangChain, Transformers Agents, and Lagent, and allows for the extension of custom tools with flexible interfaces and support for serving and remote access. Key features: - A versatile set of tools for extending LLM agents with multimodal capabilities. - Flexible tool interfaces that enable easy extension of custom tools. - Support for integrating with popular agent frameworks. - Tool serving and remote access capabilities for tools with heavy ML models. Tech stack: Python Why it matters: AgentLego provides a powerful and flexible toolkit for enhancing LLM-based agents, making it easier to build advanced conversational AI applications with a wide range of capabilities.

### Overview: BERT4Pytorch是一个超轻量级的PyTorch版本的BERT实现，旨在简化BERT的使用和理解。项目包含少量代码，大量中文注释，便于初学者快速上手。它支持加载预训练权重进行fine-tuning，并且实现了ema权重滑动平均和warmup优化器。 ### Key features: - 加载预训练BERT和RoBERTa的权重进行fine-tuning - 实现ema权重滑动平均 - 添加warmup优化器 - 实现模型权重的指数滑动平均 - 实现各种trick（如对抗训练、ema等），定义特定的layer和loss - 提供大量中文注释，便于理解BERT架构 ### Tech stack: Python ### Why it matters: BERT4Pytorch简化了BERT的使用和理解，通过浓缩的代码结构和大量的中文注释，使得初学者能够快速上手。它还支持加载预训练权重进行fine-tuning，并且实现了ema权重滑动平均和warmup优化器，使得BERT的fine-tuning过程更加高效。此外，项目还支持多种网络架构的实现，如albert、GPT、XLnet等，方便后续的扩展和改进。