Projects

### Overview: Arcana is an embeddable RAG (Relevance-Augmented Generation) library for Elixir/Phoenix. It seamlessly integrates vector search, knowledge graphs, and LLM-driven retrieval into any Phoenix app that already has an Ecto repo. This library allows developers to add vector search, knowledge graphs, and LLM-driven retrieval to their applications without the need for a separate vector database, indexing service, or orchestration layer. ### Key Features: - **One Repo**: Documents, chunks, embeddings, and the knowledge graph are stored in your existing Postgres database. - **Local-first by Default**: Embeddings run on Bumblebee, EXLA (Apple Silicon), or Torchx, and the cross-encoder reranker is also local. - **One Process Model**: Embedders and rerankers are `Nx.Serving` instances under your supervision tree, and telemetry events are `:telemetry` spans you can already consume. - **Pluggable, but Not Abstract**: Every step that can be replaced is a behavior with a single callback and a sensible default. Custom rerankers, custom searchers, and custom answerers are all 10-line modules. ### Tech Stack: Elixir ### Why It Matters: Arcana takes the BEAM (Elixir) platform seriously and leverages its strengths for RAG systems. It eliminates the need for a separate vector database, indexing service, or orchestration layer, making it a more efficient and straightforward solution for developers. The library's modular design allows for easy customization and replacement of individual components,

**Overview:** VisualTorch is a tool designed to help developers visualize Torch-based neural network architectures. It supports generating flow-style, graph-style, and LeNet-style visualizations for PyTorch Sequential and Custom models, aiming to provide a unified tracing backend and architecture-handling logic. **Key features:** - Supports flow-style, graph-style, and LeNet-style visualizations. - Traces real forward passes to build diagrams, with support for data-dependent control flow and multiple output tensors. - Provides animated GIFs for an animated reveal feature, showing layers/columns one at a time. **Tech stack:** Python **Why it matters:** VisualTorch helps developers better understand and visualize complex neural network architectures, making it easier to debug and optimize models. Its ability to show multiple output tensors and data-dependent control flow makes it particularly useful for understanding intricate models.

Overview: infinifi is a website that continuously plays gentle, lo-fi music in the background to help users relax and study. Key features: - Continuous music generation using Meta's MusicGen model - Pure frontend written in HTML/CSS/JS with no external dependencies - Automatic fade-in and fade-out effect for each music clip - Multiple audio clips saved locally for continuous playback Tech stack: Python Why it matters: The project showcases the seamless integration of AI music generation with a user-friendly frontend, providing a relaxing and efficient study environment. Repository: kennethnym/infinifi Primary language: Python GitHub stars: 342 GitHub description: infinifi plays gentle lofi music in the background indefinitely README excerpt: infinifi is a cosy little website that plays calm, gentle lo-fi music in the background to help you relax and study!

### Overview: Building and improving the performance of AI applications requires a tech-stack designed for iterative development. Palico helps streamline the process by providing flexibility, performance optimization, and seamless integration with various tools and technologies. ### Key Features: - **Build Any Application with Flexibility**: Create applications with complete control over implementation details. - **Preview Changes Locally**: Use a Playground UI to preview changes locally before deploying. - **Improve Performance with Experiments and Evaluations**: Experiment with different models, prompts, and datasets to optimize performance. - **Debug Issues with Logs and Traces**: Utilize logs and traces for efficient debugging. - **Integrate with Frontend with ClientSDKs or REST API**: Easily integrate with your frontend using ClientSDKs or REST API. - **Setup Continuous Integration and Pull-Request Previews**: Automate CI/CD processes for seamless integration. - **Manage Application from Control Panel**: Access a control panel for managing your application. - **Star this Repo for Release Notifications**: Keep track of new features by starring the repository. ### Tech Stack: TypeScript ### Why It Matters: Palico simplifies the development process by providing a tech-stack tailored for the iterative nature of LLM development. It enables you to build applications with complete flexibility, optimize performance through experiments and evaluations, debug issues effectively, integrate with your frontend, automate CI/CD, and manage your application from a control panel.

Overview: `tokenizer` is a pure Go package designed to facilitate Natural Language Processing (NLP) tasks in Go, inspired by the popular HuggingFace Tokenizers. Key features: - Supports various tokenizer models including Word level, Wordpiece, and Byte Pair Encoding (BPE). - Implements a modular design with sub-packages for normalizer, pretokenizer, tokenizer, and post-processing. - Compatible with pretrained models from Huggingface, allowing for both training and fine-tuning. Tech stack: Go. Why it matters: Bringing AI/deep-learning tools to Go developers, enabling faster software development in production environments. Repository: `sugarme/tokenizer` GitHub stars: 342 GitHub description: NLP tokenizers written in Go language README excerpt: # Tokenizer [](https://opensource.org/licenses/Apache-2.0)[](https://pkg.go.dev/github.com/sugarme/tokenizer?tab=doc)[](https://travis-ci.org/sugarme/tokenizer)[](https://goreportcard.com/report/github.com/sugarme/token

Overview: varg is an open-source TypeScript SDK designed for AI video generation, providing a single API key to generate videos with various providers like Kling, Flux, ElevenLabs, and Sora. It simplifies the process of creating AI videos by allowing developers to use declarative JSX components. Key features: - One API key for all providers - Automatic caching for instant re-rendering - Local or cloud rendering capabilities - Supports various AI providers including Kling, Flux, ElevenLabs, and Sora Tech stack: TypeScript Why it matters: varg simplifies the process of creating AI videos by providing a single API key, making it easier for developers to integrate AI generation into their projects without needing to manage multiple APIs.

Overview: This project is based on the latest [huggingface](https://github.com/huggingface) offering's [transformers](https://github.com/huggingface/transformers/releases/tag/v2.2.2) v2.2.2 code for reconfiguration. It ensures compatibility of the code for future reusability without compatibility issues. The project supports various feature extractors and provides a test set prediction code. Key features: - Support for transformer models followed by various feature extractors - Support for test set prediction code - Simplifies the original Transformers code to make it more suitable for text classification tasks - Optimizes logging terminal output for more reasonable content Tech stack: Python Why it matters: This project is designed to facilitate the use of Transformers for text classification tasks, making it easier and more efficient to implement and evaluate models. Repository: zhanlaoban/Transformers_for_Text_Classification Primary language: Python GitHub stars: 342 GitHub description: 基于Transformers的文本分类 README excerpt: # Transformers_for_Text_Classification # 基于Transformers的文本分类 基于最新的 [huggingface](https://github.com/huggingface) 出品的 [transformers](https://github.com/huggingface/transformers/releases/tag/v2.2.2) v2.2.2代码进行重构。为了保证代码日后可以直接复现而不出现兼容性问题，这里将 [transformers](https://github.com/huggingface/transformers/releases/tag

### Overview ChatKBQA is an official resource for the "ChatKBQA: A Generate-then-Retrieve Framework for Knowledge Base Question Answering with Fine-tuned Large Language Models" paper presented at ACL 2024. This framework aims to enhance the accuracy of knowledge base question answering by leveraging large language models and a generate-and-retrieve approach. The project is designed to facilitate research and practical applications in the field of knowledge base question answering. ### Key Features - **Generate-then-Retrieve Framework**: Implements a method where large language models generate answers to questions, and then retrieve the most relevant information from a knowledge base. - **Fine-tuned Large Language Models**: Utilizes pre-trained models like GPT-3 or similar to improve the accuracy of question answering. - **Freebase Knowledge Graph**: Integrates a large knowledge graph to provide context and information retrieval capabilities. ### Tech Stack - **Programming Languages**: Python - **Libraries**: PyTorch, Transformers, OpenKG, Freebase ### Why It Matters This project is significant because it provides a practical framework for enhancing the accuracy of knowledge base question answering, which is crucial in various applications such as information retrieval, educational tools, and more. The use of large language models and a knowledge graph makes it particularly effective in handling complex and context-dependent questions. ### Repository The project is hosted on GitHub at [LHRLAB/ChatKBQA](https://github.com/LHRLAB/ChatKBQA). It has been starred by 341 developers, indicating its popularity

Overview: This repository is a collection of projects completed during the "100 Days of ML Code" challenge, aimed at learning and growing in the field of Machine Learning and Artificial Intelligence. The projects cover various techniques such as data preprocessing, regression, classification, clustering, natural language processing, reinforcement learning, deep learning, dimensionality reduction, model selection, and data visualization. Key features: - Data preprocessing techniques including handling missing data, encoding categorical data, and feature scaling. - Projects on regression methods like Simple Linear Regression, Multi Linear Regression, Polynomial Regression, and Support Vector Regression. - Classification projects like Logistic Regression, K Nearest Neighbors Classification, Support Vector Machine, Kernel SVM, Naive Bayes, and Decision Tree Classification. - Clustering projects like K-Means Clustering and Hierarchical Clustering. - Association rule projects like Apriori and Eclat. - Reinforcement learning projects like Upper Confidence Bounds and Thompson Sampling. - Natural language processing projects using AWS Comprehend. - Deep learning projects including Artificial Neural Networks and Convolutional Neural Networks. - Dimensionality reduction techniques like Principal Component Analysis and Linear Discriminant Analysis. - Model selection techniques such as Grid Search and K-fold Cross Validation. - Data visualization using libraries like Matplotlib, Tableau, Power BI, and Grafana. Tech stack: The projects are implemented using Python, Jupyter Notebook, and various machine learning libraries and tools. Why it matters: This repository serves as a practical guide for developers looking to enhance their skills in machine learning and artificial intelligence. It provides a comprehensive overview of various techniques

### Overview: Taiwan Legal RAG CLI is a free, no-signup, no-API-key Taiwan legal MCP server that allows you to retrieve 2,275 million Taiwan court judgments and administrative decisions using your own AI tools. It includes a comprehensive public source-available CLI, backed by Legal Detective's 22M-judgment semantic retrieval infrastructure. ### Key Features: - **Public Source-Only CLI**: A free, no-signup, no-API-key Taiwan legal MCP server. - **Semantic Retrieval Infrastructure**: Built on Legal Detective's 22M-judgment semantic retrieval infrastructure. - **Judgment Retrieval**: Retrieves Taiwan court judgments and administrative decisions. - **Citation Check**: Verifies if cited judgments belong to the retrieved bundle. - **Case History**: Includes case history information such as the court's decision and the previous court's judgment. ### Tech Stack: Python ### Why It Matters: This tool provides a comprehensive and structured way to access and retrieve Taiwan's legal judgments and decisions, making it an essential resource for anyone looking to integrate legal information into their AI applications.

Overview: DiffusionDepth is a novel approach that transforms monocular depth estimation into a denoising diffusion process, enabling the use of generative models in sparse depth scenarios. This method learns to reverse the diffusion process of refined depth into a random depth distribution, overcoming the limitations of applying generative models to sparse ground truth depth. Key features: - Self-diffusion formulation - Iterative denoising process - Latent space encoding for depth encoder and decoder Tech stack: The repository uses Python 3.9, PyTorch 1.10, and NVIDIA CUDA 11.4 for implementation. Why it matters: DiffusionDepth addresses the challenge of applying generative models to sparse ground truth depth, making it a significant contribution to the field of monocular depth estimation. The method's ability to handle sparse data effectively opens up new possibilities for real-world applications in autonomous driving, robotics, and other areas where depth perception is crucial. GitHub stars: 341

### Overview: Agentic-Desktop-Pet 是一个结合了人工智能、记忆、情感和角色扮演功能的桌面宠物项目。它旨在为用户提供一个有趣且互动性强的虚拟伴侣体验。 ### Key features: - **记忆系统**: 使用 Cognee 知识图谱进行知识组织和持久存储，支持语义搜索和情感丰富化。 - **个人助手能力**: 实现类似 Claude Code 的功能，包括文件操作、代码执行、任务管理以及系统命令执行。 - **mod 系统**: 轻松切换和添加不同桌宠主题。 - **动态情感系统**: 基础情感状态、时间衰减、情感增强和行为影响。 - **RPG 属性系统**: 基础属性、经验等级、技能系统和好感度。 ### Tech stack: - Python 3.13+ - Godot 4.3+ - LLM API (OpenAI / DeepSeek / Claude 等) ### Why it matters: Agentic-Desktop-Pet 是一个结合了多种先进技术的桌面宠物项目，旨在提供一个有趣且互动性强的虚拟伴侣体验。通过结合记忆、情感和角色扮演功能，它为用户提供了一个全新的虚拟世界探索和互动方式。 ### Repository: - **Repository**: jihe520/Agentic-Desktop-Pet - **GitHub stars**: 340 - **Primary language**: Python - **GitHub description**: 🤖👾🐶 下一代 Agentic 桌宠 = LLM +

Overview: This repository contains code for the TriviaQA reading comprehension dataset, a large-scale dataset presented at ACL 2017. The data can be downloaded from the TriviaQA website and is licensed under the Apache 2.0 License. Key features: - Code for the paper "TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension" by Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. - Contains evaluation scripts for the dataset. - Includes instructions for running the evaluation using Python 3, with some details for compatibility with Python 2.7. - BiDAF requires Python 3, with TensorFlow verified on version 0.11. - The README provides guidance on file formats and evaluation scripts. Tech stack: Python 3 Why it matters: This dataset is significant for researchers and developers interested in reading comprehension and distant supervision techniques in natural language processing.

### Overview: Sotopia is an open-ended social learning environment designed for evaluating and facilitating social intelligence in language agents. It is presented at ICLR 2024 as a spotlight. ### Key Features: - **Open-Ended Environment**: Allows agents to interact with each other and the environment. - **Scalability**: Easily scalable to include a large number of agents and environments. - **Interactive Evaluation**: Facilitates the evaluation of social intelligence in language agents. - **Interactive Learning**: Enables agents to learn from each other and the environment. ### Tech Stack: Python ### Why it Matters: Sotopia provides a comprehensive platform for evaluating and improving social intelligence in language agents, making it a valuable tool in the field of artificial intelligence and social learning.

Overview: SynCode is a framework designed to generate Large Language Models (LLMs) with soundness and completeness guarantees, ensuring syntactic validity with respect to Context-Free Grammars (CFGs). It supports various programming languages and is notably efficient and faster than unconstrained generation. Key features: - Fast grammar-guided generation with minimal overhead. - Seamless integration with HuggingFace Language Models. - Support for built-in CFGs for Python, Go, Java, SQL, JSON, and more. - Sample usage with any existing decoding strategy. Tech stack: Python Why it matters: SynCode ensures that the generated output is syntactically valid, making it suitable for applications requiring precise and reliable language generation, such as programming and documentation. Repository: structuredllm/syncode Primary language: Python GitHub stars: 341 GitHub description: Efficient and general syntactical decoding for Large Language Models.

Overview: This repository contains a comprehensive toolkit for developers to integrate speech and text functionalities, including real-time microphone input and audio file reading. It supports both speech to text (using PocketSphinx and Baidu API) and text to speech (using pyttsx3) functionalities. Key features: - Speech to text (Pocketsphinx, Baidu API, Iflytek API) - Text to speech (pyttsx3) Tech stack: Python Why it matters: This toolkit simplifies the integration of speech and text functionalities, making it easier for developers to implement voice recognition and text synthesis in their applications. Repository: Renovamen/Speech-and-Text Primary language: Python GitHub stars: 340 GitHub description: Speech to text (PocketSphinx, Iflytex API, Baidu API) and text to speech (pyttsx3) | 语音转文字（PocketSphinx、百度 API、科大讯飞 API）和文字转语音（pyttsx3） README excerpt: # Speech-and-Text 语音转文字（支持实时麦克风输入和从音频文件读入）： - 百度 API - 科大讯飞 API - SpeechRecognition (CMU PocketSphinx) 文字转语音： - pyttsx3   ## Environment - Python 3.6.7 - MacOS（以下环境配置方式均基于Mac系统，其他系统的配置方式可能会有一些不同）   ## Speech to Text ### 百度 在 https://cloud.baidu.com/product/speech

### Overview: TLive-Omni is an omni-modal understanding model designed for e-commerce live-streaming, capable of mapping image, video, audio, and text into a unified text-output interface. Built on a Qwen3.5 backbone with an audio encoder, it supports 256K tokens of context and is trained using a three-stage SFT recipe followed by Faithful-RFT reinforcement fine-tuning. ### Key features: - **Timestamped Per-vGrid layout**: Organizes audio and video tokens into timestamped grids, maintaining adjacent audio segments with visual content for fine-grained temporal alignment. - **Three-stage SFT recipe**: Progressively trains from audio-language alignment to full multimodal SFT, developing live-commerce understanding from omni-modal perception to instruction-following responses. - **Faithful-RFT**: Reinforcement fine-tuning stage for faithful and real-time live-stream demands, suppressing explicit reasoning traces and optimizing answer quality for live-commerce tasks. - **Rich atomic capabilities**: Covers speech recognition, speaker analysis, product visual grounding, text recognition, temporal grounding, video dense caption, and omni-modal QA, supported by a compact data production engine. - **Strong live-commerce performance with competitive generalization**: Demonstrates strong results across live-commerce audio, image, and video tasks, and excels on general benchmarks. ### Tech stack: Qwen3.5 backbone with a grafted audio encoder. ### Why it matters: TLive-Omni is designed to enhance the understanding and interaction capabilities of e-commerce live-streaming, providing a unified interface for multim

Overview: MedQA-ChatGLM is a medical QA model trained on real medical dialogue data using LoRA, P-Tuning V2, Freeze, and RLHF techniques. It leverages various datasets including cMedQA2, ChatMed, ChatDoctor, and others to enhance its performance in medical Q&A tasks. Key features: - Fine-tuned on multiple datasets including cMedQA2, ChatMed, ChatDoctor, and others. - Uses LoRA, P-Tuning V2, Freeze, and RLHF for micro-tuning. - Employs various datasets to improve its performance in medical Q&A. Tech stack: Python Why it matters: This model is designed to enhance its performance in medical Q&A tasks by leveraging real medical dialogue data and various micro-tuning techniques. It is particularly useful for applications requiring high accuracy in medical conversations. Repository: WangRongsheng/MedQA-ChatGLM GitHub stars: 340 GitHub description: 🛰️ 基于真实医疗对话数据在ChatGLM上进行LoRA、P-Tuning V2、Freeze、RLHF等微调，我们的眼光不止于医疗问答 GitHub: https://github.com/WangRongsheng/MedQA-ChatGLM

Overview: Advising App® by Canyon GBS® is an AI-powered conversational assistant designed specifically for colleges and universities to enhance student engagement and drive institutional success. Key features: - Student Messaging - AI Chat - Guided Outreach Tech stack: PHP Why it matters: Advising App® simplifies communication, provides AI support, and integrates with campus systems for seamless student engagement and relationship management. Repository: canyongbs/advisingapp

Overview: MemN2N is an implementation of End-To-End Memory Networks using Tensorflow, designed for tasks from the bAbl dataset. Key features: - Uses a sklearn-like interface - Supports multiple tasks - Implements stochastic gradient descent with an annealed learning rate schedule Tech stack: Python with Tensorflow Why it matters: MemN2N is a significant contribution to the field of neural network research, offering a practical solution for end-to-end memory network tasks. Repository: domluna/memn2n Primary language: Python GitHub stars: 340 GitHub description: End-To-End Memory Network using Tensorflow README excerpt: Implementation of [End-To-End Memory Networks](http://arxiv.org/abs/1503.08895) with sklearn-like interface using Tensorflow. Tasks are from the [bAbl](http://arxiv.org/abs/1502.05698) dataset.