Projects

### Overview: 本项目自动抓取并索引科学空间的文章元数据，按研究主题进行规则分类，方便在 GitHub 上快速浏览并跳转到原文。 ### Key features: - 自动抓取和索引科学空间文章 - 按研究主题分类文章 - 提供规则分类的 Markdown 文件 - 便于在 GitHub 上快速浏览和跳转 ### Tech stack: HTML ### Why it matters: 通过提供忠实于原文的 Markdown 文件，帮助开发者更高效地查找和阅读科学空间的文章，从而更好地理解和应用相关知识。

Overview: 深度研搜对话式多智能体研究系统是适合系统学习 DeepAgents 的实战项目，配套有完整代码、教程和教程分支，适合企业级 AI 工程落地。 Key features: - 多智能体架构，主智能体调度专家助手分工。 - 多来源检索，包括互联网搜索、数据库查询和私有知识库。 - 从检索到交付的完整链路，包括文件工具处理和前端展示。 - 会话级上下文隔离，工具调用和结果展示通过 WebSocket 推送。 - 工程化前后端结构，基于 FastAPI、WebSocket、DeepAgents 和 React 组织任务接口。 Tech stack: Python Why it matters: 适合系统学习大模型应用、多智能体 Agent 和企业级 AI 工程落地。

Overview: GradsFlow is an open-source AutoML library built on PyTorch, designed to democratize AI by providing easy-to-use model training capabilities directly on a laptop or a remote cluster. Key features: - Model Training API - PyTorch Lightning Flash support - Zero-code AutoML with AutoTasks - Custom model training and hyperparameter optimization with Model and Tuner Tech stack: Python Why it matters: GradsFlow aims to make AI accessible to everyone by simplifying the process of building and training deep learning models, making it easier to democratize AI and democratize the use of AI in various applications.

### Overview: AudioLDM training, finetuning, evaluation, and inference are essential components of the AudioLDM project. This repository provides a comprehensive set of tools and scripts for training, fine-tuning, evaluating, and deploying the AudioLDM model. ### Key Features: - **Model Training**: The repository includes scripts for training the AudioLDM model from scratch and fine-tuning pre-trained checkpoints. - **Dataset Preparation**: A preprocessed dataset is provided for training and evaluation. - **Inference**: The model can be used for generating audio outputs directly. - **Evaluation**: Automatic evaluation is performed on the generated audio outputs to assess model performance. ### Tech Stack: - Python ### Why it Matters: The AudioLDM project is crucial for generating high-quality audio content from text inputs, making it valuable for applications such as automatic speech recognition, music generation, and more. ### Repository: - **GitHub Repository**: haoheliu/AudioLDM-training-finetuning - **Stars**: 306 - **Description**: AudioLDM training, finetuning, evaluation, and inference. ### Primary Language: - Python ### GitHub Description: - AudioLDM training, finetuning, evaluation and inference. ### How to Use: 1. **Prepare Python Running Environment**: - Install a conda environment. - Clone the repository and set up the environment. - Download and place the required checkpoints and dataset into the specified directories. 2. **Play around with the Code**: - Train the AudioLDM model.

Overview: The pinyin project is a high-performance Chinese to pinyin converter tool written in Java. Key features: - Converts text to pinyin - Supports conversion of long texts - Handles multiple tones - Provides various pinyin annotation styles - Supports Chinese character segmentation - Offers a custom pinyin dictionary - Checks for homophones - Handles homophones Tech stack: Java Why it matters: The project aims to provide a simple and efficient pinyin tool for Java developers, focusing on performance and ease of use. Repository: houbb/pinyin Primary language: Java GitHub stars: 306

Overview: Langtorch is a Java library designed for building composable Large Language Models (LLMs) applications, providing a robust and efficient solution for developers. Key features: - Composable LLM applications - Workflow automation - Easy integration with existing Java projects Tech stack: Java Why it matters: Langtorch simplifies the development of LLM applications, making it easier for developers to integrate advanced AI capabilities into their projects without the need for extensive expertise in AI or programming. Repository: li2109/langtorch Primary language: Java GitHub stars: 306 GitHub description: 🔥 Building composable LLM applications & workflow with Java.

### Overview: Transformer in Transformer is an implementation of a transformer model within a transformer model, designed for image classification tasks. This innovative approach combines pixel-level attention with patch-level attention, enabling precise classification of images. ### Key features: - Pixel level attention - Patch level attention - PyTorch framework - Implementation of Transformer in Transformer model ### Tech stack: PyTorch ### Why it matters: This project showcases the potential of combining two different types of attention mechanisms within a transformer architecture, potentially leading to improved performance in image classification tasks.

Overview: PoisonedRAG is an official repository of a paper presenting a knowledge corruption attack on Retrieval-Augmented Generation (RAG) systems. The paper was accepted to USENIX Security 2025 and is focused on evaluating and improving the security of large language models. Key features: - Knowledge Corruption Attacks - Retrieval-Augmented Generation (RAG) System - Official Repository of [USENIX Security 2025](https://www.usenix.org/conference/usenixsecurity25) Tech stack: Python Why it matters: This repository provides a comprehensive framework for evaluating the security of RAG systems, highlighting the importance of ensuring the integrity and reliability of large language models in the face of potential attacks.

### Overview: This skill provides a comprehensive workflow for transforming AI-generated images and videos into high-quality game assets, including sprites, icons, backgrounds, video-to-frames animation, cinematic effects, and runtime checks. It is designed to ensure assets can survive the rigors of a real game loop. ### Key Features: - **Sprites**: Plan and generate sprites for action RPGs, ensuring they are transparent and do not jitter. - **Icons**: Create and clean up icons, focusing on readability and stability. - **Backgrounds**: Design backgrounds that fit mobile GPU texture limits. - **Cinematic Effects**: Convert video to frames and add code-driven hit logic for cinematic effects. - **Runtime Checks**: Ensure assets are ready for mobile runtime shipping. ### Tech Stack: - Python ### Why It Matters: Most AI game art demos stop at a beautiful image. Games need stricter quality standards, including transparent sprites, stable animations, background fits, cinematic effects, and runtime checks. The skill ensures assets can withstand the rigors of a real game loop, making them suitable for production use. ### Repository Layout: - **README.md**: Provides an overview and installation instructions. - **references/**: Contains detailed documentation and examples for each asset type. - **scripts/**: Contains scripts for generating and cleaning assets. - **examples/**: Contains examples and scripts for various tasks. - **SKILL.md**: Describes the skill's capabilities and recommended workflows. - **scripts/**: Contains scripts for generating and cleaning assets. - **examples/**: Contains examples and scripts for various tasks. - **scripts

### Overview: AX-Samples 是由爱芯元智主导开发的项目，旨在提供常见的深度学习开源算法在爱芯元智的AI SoC上的示例代码，方便社区开发者快速评估和适配。 ### Key features: - 示例代码覆盖了常见的深度学习开源算法，包括分类、检测、分割和姿态估计等。 - 支持多种芯片平台，包括终端计算芯片（如AX650A、AX650N、AX637等）、车载芯片（如M57）和边缘计算芯片（如AX8850N、AX8850、AX8860）。 - 提供详细的编译文档、示例结果图和开发板图片，方便开发者快速上手。 ### Tech stack: CMake ### Why it matters: AX-Samples为开发者提供了快速评估和适配AI SoC的工具，简化了深度学习算法在硬件上的实现和部署过程，有助于加速AI技术在实际应用中的落地。

Overview: HFlow is an open-source SDK designed to help robotics teams verify the quality of their data used for AI model training. It streamlines the process of collecting, processing, and storing multimodal data, making it easier for teams of any size to manage and audit their data. Key features: - Built-in checks for data quality - Support for Python transforms, checks, labels, and enrichments - Orchestration, storage, versioning, and curation of data pipelines - Timestamped provenance and metadata tracking - Graph visualization of the data pipeline - Queryable catalog of quality evidence Tech stack: Python Why it matters: Processing data is a significant bottleneck in robotics, and HFlow provides a solution to make data tooling accessible to teams of any size. By automating the process of verifying data quality, HFlow helps teams ensure that their AI models are trained on high-quality data, leading to better outcomes and more reliable models. Repository: Hebbian-Robotics/hflow Primary language: Python GitHub stars: 302 GitHub description: SDK for scalable multimodal data pipelines in robotics and physical AI README excerpt: Hebbian Robotics (YC S26) is building HFlow, an open source SDK for scalable multimodal data pipelines in robotics and physical AI. It makes data tooling and practices typically developed inside large robotics teams accessible to teams of any size. Processing data is a major bottleneck in robotics, and HFlow provides a solution to make data tooling accessible to teams of any size. By automating the

Overview: HeyClaude is a curated registry and distribution surface for Claude and AI-workflow assets, including agents, MCP servers, skills, commands, hooks, rules, guides, templates, and more. Key features: - Curated Claude workflow infrastructure - 1388+ file-backed entries - Human-reviewed and machine-readable registry - API access to MCP package and remote MCP endpoint - Raycast extension and RSS feed integration - Jobs board for developers Tech stack: JSONb Why it matters: HeyClaude serves as a valuable resource for Claude developers, providing a curated directory of assets that are both useful as awesome-list catalogs and machine-readable registries.

### Overview: This project aims to simulate the learning process of human brain cells, specifically focusing on the cerebellar lateral cell (CL1) neurons. Using a Doom environment, it trains a policy to control the firing of these neurons, thereby enabling the cells to learn through their interactions with the environment. ### Key Features: - **Training Environment**: The project uses a Doom environment to simulate the learning process of human brain cells. - **Neuron Model**: The CL1 neurons are trained to learn through their interactions with the environment. - **Policy Gradient Method**: The policy gradient method (PPO) is used to train the policy, with the encoder/PPO learning from the CL1 neurons. ### Tech Stack: - Python ### Why it Matters: This project provides a detailed simulation of how neurons learn through their interactions with the environment, specifically focusing on the cerebellar lateral cell (CL1) neurons. It demonstrates the importance of learning through experience and the role of the policy gradient method in training policies that can mimic biological learning processes.

### Overview: SoulX-Transcriber is an end-to-end framework designed for multi-speaker transcription, jointly modeling who spoke, when, and what, achieving superior performance on the AISHELL-4 and AliMeeting benchmarks. ### Key Features: - **Speaker Attribution**: The model learns to attribute each utterance to a specific speaker. - **Timestamped Segmentation**: It produces accurate timestamps for segmented audio. - **Transcription**: It generates coherent transcripts for overlapping and fast-turn conversations. ### Tech Stack: - Python ### Why It Matters: SoulX-Transcriber provides a unified framework for diarization and recognition, producing structured outputs that are highly aligned with speaker labels and transcriptions, making it particularly effective for multi-speaker dialogue scenarios. ### Repository: The repository is hosted on GitHub at [Soul-AILab/SoulX-Transcriber](https://github.com/Soul-AILab/SoulX-Transcriber). ### GitHub Stars: The repository has received 303 stars from the developer community. ### Primary Language: The primary language used in the project is Python. ### Demo Video: A demo page is available at [https://soul-ailab.github.io/soulx-transcriber/](https://soul-ailab.github.io/soulx-transcriber/). ### Performance Overview: SoulX-Transcriber achieves state-of-the-art performance on the AISHELL-4 and AliMeeting benchmarks, demonstrating its effectiveness in multi-speaker transcription scenarios. ### Speaker-Aware Multi-Stage Training

Overview: This is a PyTorch implementation of the ACL 2019 paper "Simple and Effective Text Matching with Richer Alignment Features". The model aims to explore what is sufficient for strong performance in text matching tasks. Key features: - Richer alignment features - Original point-wise features - Previous aligned features - Contextual features Tech stack: PyTorch Why it matters: RE2 achieves performance on par with the state of the art on four benchmark datasets: SNLI, SciTail, Quora, and WikiQA, across tasks of natural language inference, paraphrase identification, and answer selection with no or few task-specific adaptations. It has at least 6 times faster inference speed compared to similarly performed models. Repository: alibaba-edu/simple-effective-text-matching-pytorch GitHub stars: 305

### Overview: Text2Text Language Modeling Toolkit is a comprehensive toolkit designed for developers to create, train, and utilize language models. It provides a wide range of functionalities including assistant, tokenization, embedding, TF-IDF, BM25, index, translation, and data augmentation. ### Key Features: - **Assistant**: A free private ChatGPT-like alternative. - **Tokenization**: Converts text into a sequence of tokens. - **Embedding**: Converts text into numerical vectors. - **TF-IDF**: Measures the importance of terms in a document. - **BM25**: A ranking algorithm for retrieving documents based on relevance. - **Index**: Provides an index for efficient information retrieval. - **Translation**: Translates text from one language to another. - **Data Augmentation**: Enhances the quality of training data through back-translation. ### Tech Stack: Python ### Why It Matters: Text2Text is essential for developers looking to integrate AI-powered text generation into their applications. Its comprehensive set of features makes it a versatile tool for various language modeling tasks, from simple text processing to complex language translation and data augmentation. ### GitHub Stars: 305 ### Installation Requirements: ```bash pip install -qq -U text2text ``` ### Examples: - **Assistant**: `t2t.Assistant().transform("Describe Text2Text in a few words: ")` → "Text2Text is an AI-powered text generation tool that creates coherent and continuous text based on prompts." - **Tokenization**: `t2t.Tokenizer().transform(["Hello

Overview: SciRS2 is a Pure Rust scientific computing and AI/ML library that offers SciPy-compatible APIs, providing 10-100x performance gains through SIMD optimization and no system dependencies. Key features: - Pure Rust implementation - 10-100x performance improvements - Memory safety - Cross-platform compatibility - Production-ready quality Tech stack: Pure Rust Why it matters: SciRS2 is a comprehensive scientific computing and AI/ML infrastructure in Pure Rust, providing SciPy-compatible APIs while leveraging Rust's performance, safety, and concurrency features. Unlike traditional scientific libraries, SciRS2 is 100% Pure Rust by default with no C/C++/Fortran dependencies required, making installation effortless and ensuring cross-platform compatibility. Repository: cool-japan/scirs GitHub stars: 302

### Overview: MedMCQA is a large-scale, Multiple-Choice Question Answering (MCQA) dataset designed to address real-world medical entrance exam questions. The dataset consists of over 194,000 high-quality AIIMS & NEET PG entrance exam MCQs covering 2.4k healthcare topics and 21 medical subjects. Each sample includes a question, correct answer(s), and other options that require a deeper language understanding. ### Key Features: - **High-Quality Questions:** Over 194,000 questions collected from AIIMS & NEET PG entrance exams. - **Wide Range of Topics:** Covers 2.4k healthcare topics and 21 medical subjects. - **Deep Language Understanding:** Tests reasoning abilities across a wide range of medical subjects and topics. - **Open-Source Dataset:** Provides an open-source dataset for the Natural Language Processing community. ### Tech Stack: - **Repository:** GitHub - **Primary Language:** Jupyter Notebook ### Why It Matters: This dataset is crucial for researchers and developers in the Natural Language Processing (NLP) community, as it provides a comprehensive and diverse dataset for improving QA systems. It can be used to train and evaluate models that understand medical language and reasoning abilities, making it valuable for applications in healthcare and medical education.

### Overview: Reviva 是一个集成了本地资料库、Wiki 知识库、AI Agent 和创作工具的本地学习工作台。它旨在帮助用户高效地管理和学习资料，通过对话、笔记、复习和创作输出，实现知识的持续更新和沉淀。 ### Key features: - 📚 资料可以被导入、解析、检索和持续维护 - 💬 对话可以分组、分支，并保留上下文与引用来源 - 🧠 可在同一个会话中切换Agent，共享上下文 - ✨ 一份资料可以继续生成测验、闪卡、导图、图表、PPT、播客或研究报告 - 🔁 输出可以保存到笔记、Wiki 和本地目录，形成学习闭环 ### Tech stack: HTML ### Why it matters: Reviva 是一个集成了本地资料库、Wiki 知识库、AI Agent 和创作工具的本地学习工作台。它旨在帮助用户高效地管理和学习资料，通过对话、笔记、复习和创作输出，实现知识的持续更新和沉淀。

Overview: DiffusioNeRF is a novel approach that integrates Neural Radiance Fields with Denoising Diffusion Models to regularize and improve the quality of NeRF-based rendering. Key features: - Regularizes Neural Radiance Fields (NeRF) to enhance rendering quality - Utilizes Denoising Diffusion Models for efficient training and inference - Supports both RGBD patch and NeRF-based datasets Tech stack: Python Why it matters: This method addresses the limitations of current NeRF-based rendering by providing a more stable and visually appealing output, making it particularly useful for applications requiring high-quality visualizations and realistic scenes.