Projects

Overview: Chevron is a powerful and highly functional startpage integrated with chatGPT, featuring a minimalistic and animated design. It is available in static, hosted, and GitHub Pages options. Key features: - Autosuggestions - History - Macros and commands - Macros menu - Hotkeys - ChatGPT integration - Currency converter - Calculator - Animated and minimalist design Tech stack: JavaScript Why it matters: Chevron offers a seamless integration of chatGPT, a minimalist design, and powerful features, making it a valuable tool for developers looking for a modern and functional startpage. Repository: kholmogorov27/chevron Primary language: JavaScript GitHub stars: 404 GitHub description: Powerful and highly functional startpage integrated with chatGPT and hidden under the super minimalistic and animated design (static/hosted/github pages) README excerpt: # Chevron [](/LICENSE.md)  <p align=center> <img width="80%" src="https://i.imgur.com/Wa7HcuW.png"> </p> <p align=center> Сhevron is a <i>powerful</i> and highly <i>functional</i> startpage integrated with chatGPT<br/> and hidden under the super <i>minimalistic</i> and <i>animated</i> design

Overview: ReazonSpeech is a massive open Japanese speech corpus repository, providing access to the world's largest open Japanese speech corpus. Key features: - ReazonSpeech: Building the world’s largest open Japanese speech corpus. - AVista 🐦‍🔥: Project for noise-robust multimodal speech recognition for human-robot interaction and beyond. - `reazonspeech.k2.asr <pkg/k2-asr>`_: Next-gen Kaldi model with fast and accurate speech recognition. - `reazonspeech.nemo.asr <pkg/nemo-asr>`_: Fast, accurate speech recognition based on FastConformer-RNNT. - `reazonspeech.espnet.asr <pkg/espnet-asr>`_: Speech recognition with a Conformer-Transducer model. - `reazonspeech.avsr <pkg/avsr>`_: Audio-Visual Speech models for AVista 🐦‍🔥. - `reazonspeech.evaluation <pkg/evaluation>`_: Provides tools to evaluate ReazonSpeech models and other speech recognition models. - `reazonspeech.espnet.oneseg <pkg/espnet-oneseg>`_: Tools to analyze Japanese "one-segment" TV stream. Tech stack: Python Why it matters: ReazonSpeech is a crucial resource for researchers and developers working on speech recognition and multimodal interaction, providing a comprehensive and accessible corpus for advanced research and development.

Overview: TigerLab is an open-source toolkit designed to facilitate the development of trustworthy Large Language Models (LLMs). It includes features such as TigerRAG for enhanced retrieval capabilities, TigerTune for fine-tuning, and TigerArmor for AI safety evaluation. Key features: - TigerRAG: Uses embeddings-based retrieval, retrieval-augmented generation, and generation-augmented retrieval to improve query fulfillment. - TigerTune: A Python SDK for fine-tuning, inference, and evaluation of Text Generation and Text Classification models. - TigerDA: A data augmentation toolkit for generating augmented data with fine-tuned models. - TigerArmor: An AI safety toolkit providing metrics, datasets, and evaluation tools for measuring AI safety in LLMs. Tech stack: Python, Jupyter Notebook, FAISS, BERT, text-davinci-003, Llama2, DistilBERT, GPT2, OpenAI API token. Why it matters: This toolkit is essential for developers aiming to create LLMs that are both effective and safe, ensuring they can handle complex queries and maintain ethical standards in their applications.

Overview: This is a demo showcasing the V-RAG (Vision RAG) architecture, utilizing a vision language model (VLM) to embed pages of PDF files directly, bypassing the tedious chunking process. Key features: - V-RAG architecture - Utilizes a vision language model (VLM) for embedding PDF pages directly - ColPali as the VLM in the demo - QDrant as the vector database - GPT4o or GPT4o-mini for generating responses based on queries and images Tech stack: TypeScript Why it matters: This demo showcases an innovative approach to document search and retrieval, leveraging cutting-edge technology to improve the efficiency and accuracy of document-based information retrieval.

### Overview: Freeze-Omni is a speech-to-speech dialogue model built on a frozen LLM, designed to maintain the original intelligence while achieving low latency. It includes a speech encoder and decoder, three key strategies for efficient streaming, and a model server strategy for real-time interaction. ### Key Features: - **Frozen LLM Backbone**: Built on a frozen LLM to preserve intelligence. - **Chunk-wise Streaming Input**: Supports fast response to input with a 3-stage training strategy. - **AR-based Speech Output**: Uses a single codebook AR decoder for low-latency speech synthesis. - **State Prediction**: Adds a classification layer to predict different dialogue states. ### Tech Stack: Python ### Why it Matters: Freeze-Omni is designed to maintain the intelligence of the LLM while achieving low latency, making it suitable for real-time applications and interactive dialogue systems.

### Overview: Airy is an open-source, fully-featured, production-ready data platform designed to build streaming apps with real-time data. It supports a wide range of sources, including Facebook, WhatsApp, Google's Business Messages, SMS, and custom connectors, and simplifies complex data ingestion and stream data to applications. ### Key features: - **Connectors**: Pre-built and easily configurable connectors for various sources including Facebook, WhatsApp, Google's Business Messages, SMS, and custom connectors. - **APIs**: APIs to access data with blazing fast HTTP endpoints. - **WebSockets**: Power real-time applications using WebSockets. ### Tech stack: Java ### Why it matters: Airy simplifies the process of ingesting and processing real-time data, making it easier to build and deploy ML and AI applications. Its connectors and APIs streamline the integration with various sources, making it a versatile tool for developers looking to build real-time data pipelines.

### Overview: SpikeX is a collection of pipes designed to be easily integrated into a spaCy pipeline, aimed at facilitating the creation of knowledge extraction tools with minimal effort. ### Key features: - **WikiPageX**: Links Wikipedia pages to chunks in text. - **ClusterX**: Picks noun chunks in a text and clusters them based on the Ball Mapper algorithm. - **AbbrX**: Detects abbreviations and acronyms, linking them to their long form. - **LabelX**: Takes labelings of pattern matching expressions and catches them in a text. - **PhraseX**: Creates a Doc's underscore extension based on custom attribute names and phrase patterns. - **SentX**: Detects sentences in a text, using Splitta with refinements. ### Tech stack: Python ### Why it matters: SpikeX simplifies the process of creating knowledge extraction tools by providing pre-built pipes that can be easily integrated into a spaCy pipeline, reducing the need for extensive coding and manual annotation.

### Overview: This project presents an innovative approach to zero-shot appearance transfer using cross-image attention, leveraging semantic understanding from generative models to seamlessly merge the structure of one image with the appearance of another. ### Key features: - Cross-image attention mechanism for zero-shot appearance transfer - Semantic correspondence between images - Zero-shot capability without optimization or training - Three mechanisms to improve output image quality ### Tech stack: Python ### Why it matters: This method demonstrates significant potential for applications in image synthesis, zero-shot transfer, and semantic understanding across different domains, offering a novel solution to the challenges in image appearance transfer and zero-shot learning.

Overview: The IQ Series is a comprehensive learning experience for Microsoft IQ, a unified intelligence layer for the enterprise. It includes video episodes, Jupyter notebooks, and Azure deployment templates. Key features: - Foundry IQ: Unify business semantics across data, models, and systems. - Work IQ: Unlock your organization's intelligence for every agent. - Fabric IQ: Live grounding in the public web. - Web IQ: Retrieve current, cited information beyond enterprise data. Tech stack: Jupyter Notebook Why it matters: The IQ Series provides a hands-on approach to learning Microsoft IQ, enabling AI agents to reason, retrieve, and act with deep business context, surpassing traditional RAG for true enterprise intelligence. Repository: microsoft/iq-series GitHub stars: 397 GitHub description: The IQ Series is a hands-on learning experience for Microsoft IQ: Microsoft's unified intelligence layer for the enterprise, spanning Foundry IQ, Work IQ, and Fabric IQ. The series includes video episodes, Jupyter notebooks, and Azure deployment templates.

### Overview: This repository collects and shares a wide range of resources and techniques for accelerating generation tasks, focusing on both training-free and token-wise acceleration methods. It aims to help developers and researchers improve the efficiency and speed of image and video generation models. ### Key features: - **Fast Sampling**: Techniques for accelerating sampling processes in diffusion models. - **Pruning**: Methods for reducing the size of neural networks without significant loss in performance. - **Quantization**: Techniques for reducing the precision of model parameters to speed up inference. - **Distillation**: Methods for transferring knowledge from large models to smaller ones to improve efficiency. - **Cache Mechanism**: Improvements in memory management to speed up model execution. - **Efficient Attention**: Techniques for optimizing attention mechanisms to reduce computational cost. - **Dynamic Neural Networks**: Adaptation of neural networks to optimize performance under varying conditions. - **Deployment Optimization**: Strategies for making models more efficient for deployment in real-world applications. ### Tech stack: The repository uses Python and TensorFlow for implementing and testing the techniques. ### Why it matters: This repository provides a comprehensive collection of cutting-edge methods for accelerating generation tasks, making it an essential resource for developers and researchers looking to improve the performance and efficiency of their models.

### Overview: simpleT5 is a Python library that simplifies the process of training T5 models. It leverages PyTorch Lightning and Transformers to enable quick and efficient training of various T5 models, including T5, mT5, byT5, and CodeT5. ### Key Features: - **Training in 3 Lines of Code**: Easily train T5 models using just 3 lines of code. - **Support for Multiple Models**: Works with T5, mT5, byT5, and CodeT5 models. - **Lightweight and Efficient**: Built on PyTorch Lightning and Transformers, making it lightweight and efficient. ### Tech Stack: - **PyTorch Lightning**: Framework for building and training machine learning models. - **Transformers**: Library for natural language processing, providing pre-trained models for various NLP tasks. ### Why It Matters: - **Ease of Use**: Simplifies the process of training complex NLP models, making it accessible to developers and researchers. - **Versatility**: Supports multiple T5 models, making it suitable for a wide range of applications in NLP. - **Efficiency**: Reduces the time and effort required to train and deploy models, enabling faster innovation and development. ### Repository: - **GitHub Repository**: Shivanandroy/simpleT5 - **Stars**: 402 - **Description**: simpleT5 is built on top of PyTorch-lightning and Transformers, allowing quick training of T5 models. ### Installation: - **Prerequisites**: Ensure you have Python installed and

### Overview: DeepQA is a deep learning library built on top of Keras and TensorFlow, designed specifically for high-level natural language processing tasks, particularly question answering. It simplifies the process of building and training models for various NLP tasks, making it easier to experiment with different configurations and models. ### Key Features: - **Ease of Use**: DeepQA provides a consistent and easy-to-use API for building NLP models, reducing the complexity of implementing sequence processing and attention mechanisms. - **Flexibility**: It allows for easy customization of model parameters, such as word representations and sentence encoders, by specifying them in a JSON file, making it simple to switch between different datasets and models. - **State-of-the-Art Models**: DeepQA includes a variety of advanced models tailored for question answering, such as BERT-based models, which are efficient and effective for this task. ### Tech Stack: - **Python 3.5+** - **Keras** - **TensorFlow** ### Why It Matters: - **Ease of Use**: Simplifies the implementation of complex NLP tasks, making it accessible to developers without deep expertise in deep learning. - **Flexibility**: Allows for easy experimentation with different model configurations, reducing the learning curve for new models. - **State-of-the-Art**: Provides access to advanced models, such as BERT, which are highly effective for question answering tasks. ### Repository: allenai/deep_qa - **Stars**: 402 - **Description**: A deep NLP library based on Keras / TensorFlow, focused on

Overview: Libro 是一个基于 Jupyter Notebook 的灵活定制、轻松集成的开发工具包。它提供完整的 Notebook 能力和丰富的周边功能，支持内核级别的扩展和定制。 Key features: - 🚀 提供完整的 Notebook 能力和丰富的周边功能，基于现有组合快速上手。 - 🌱 提供内核级别的扩展能力，所有层次的内容均允许二次开发定制。 - 🔮 定义大模型工作流，内置大模型交互和辅助开发能力。 Tech stack: Jupyter Notebook Why it matters: Libro 是一个高效、灵活且易于集成的开发工具，适用于数据科学家、工程师和研究人员，帮助他们快速构建和测试模型，提高工作效率和创新能力。 Repository: apt-oss/libro Primary language: Jupyter Notebook GitHub stars: 402 GitHub description: A Notebook with Flexible Customization and Easy Integration. README excerpt: <p align="center"><img src="https://raw.githubusercontent.com/wiki/difizen/libro/assets/libro-text.svg" width="160" /></p> <p align="center"><strong>libro：灵活定制、轻松集成的 Notebook 产品方案</strong></p> <p align="center"> <a href="https://github.com/difizen/libro/actions/workflows/ci.yml"><img src="https://img.shields.io/github/actions/workflow/status/difizen/libro/ci.yml?branch=main&style=for-the-badge&logo=github" alt="Code: CI

Overview: DiffLinker is an Equivariant 3D-Conditional Diffusion Model designed for the design of molecular linkers. It can seamlessly integrate any number of disconnected fragments into a molecule, without requiring information about attachment atoms or linker sizes. Key features: - Integrates any number of disconnected fragments into a molecule. - Does not require information on attachment atoms or linker size. - Can be conditioned on protein pockets. Tech stack: Python Why it matters: DiffLinker offers a versatile solution for designing linkers in molecular biology, particularly in the context of protein engineering and drug discovery, where precise control over linker placement and composition is crucial. GitHub stars: 402

Overview: This project is a set of Jupyter notebooks designed to teach developers how to effectively use ChatGPT for prompt engineering, providing a comprehensive guide on how to leverage the capabilities of large language models to build applications. Key features: - Summarizing (e.g., summarizing user reviews for brevity) - Inferring (e.g., sentiment classification, topic extraction) - Transforming text (e.g., translation, spelling & grammar correction) - Expanding (e.g., automatically writing emails) Tech stack: Jupyter Notebook Why it matters: This project is essential for developers looking to integrate ChatGPT into their applications, offering practical examples and hands-on experience to help them build powerful and efficient solutions.

Overview: This repository contains a comprehensive Chinese text classification project using PyTorch, including various neural network models and pre-trained models like BERT and ERNIE. It is designed to be a ready-to-use solution for text classification tasks. Key features: - TextCNN - TextRNN - TextRNN_Att - TextRCNN - FastText - DPCNN - Transformer - Bert - Ernie Tech stack: PyTorch Why it matters: This project provides a complete solution for Chinese text classification tasks, including state-of-the-art neural network models and pre-trained models, making it a valuable resource for developers looking to implement text classification in their projects. Repository: JackHCC/Chinese-Text-Classification-PyTorch Primary language: Python GitHub stars: 401 GitHub description: 中文文本分类任务，基于PyTorch实现（TextCNN，TextRNN，FastText，TextRCNN，BiLSTM_Attention, DPCNN, Transformer，Bert，ERNIE），开箱即用！

### Overview Few-NERD is a large-scale, fine-grained manually annotated named entity recognition dataset. It contains 8 coarse-grained types, 66 fine-grained types, 188,200 sentences, 491,711 entities, and 4,601,223 tokens. The dataset is split into three training modes: supervised (`supervised`), few-shot (`inter`), and intra (`intra`). Each mode contains three files: `train.txt`, `dev.txt`, and `test.txt`. The `supervised` datasets are randomly split, while the `inter` datasets are split within coarse type, and the `intra` datasets are split by coarse type. ### Key Features - **Coarse-grained Types**: 8 coarse-grained types. - **Fine-grained Types**: 66 fine-grained types. - **Sentences**: 188,200 sentences. - **Entities**: 491,711 entities. - **Tokens**: 4,601,223 tokens. ### Tech Stack - **Python**: The primary language used to develop the dataset and associated tools. ### Why It Matters Few-NERD provides a comprehensive dataset for evaluating and improving the performance of named entity recognition models, particularly in the context of few-shot learning. The dataset's rich annotation and diverse sentence types make it valuable for researchers and practitioners aiming to enhance their models' ability to recognize named entities in limited data scenarios. ### Repository

Overview: OSWorld 2.0 is a benchmarking project designed to evaluate computer use agents on long-horizon real-world tasks. Key features: - Real-world tasks simulation - Benchmarking of various agents - Detailed performance metrics tracking Tech stack: Python Why it matters: OSWorld 2.0 provides a comprehensive platform for evaluating the capabilities of different computer use agents, offering valuable insights into their performance across various tasks. Repository: xlang-ai/OSWorld-V2 Primary language: Python GitHub stars: 391 GitHub description: OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks

### Overview: This project combines a Convolutional Neural Network (CNN) with a Linear Support Vector Machine (SVM) for image classification. The results show that the CNN-SVM model achieved a test accuracy of approximately 99.04% on the MNIST dataset, significantly higher than the CNN-Softmax model's 99.23% accuracy. ### Key Features: - **Model Combination**: Combines CNN and SVM for image classification. - **Dataset**: MNIST dataset (8,000 images of handwritten digits). - **Accuracy**: CNN-SVM achieved ~99.04% accuracy, while CNN-Softmax reached ~99.23%. - **Preprocessing**: Empirical data indicates that better results can be achieved with improved data preprocessing and a more sophisticated base CNN model. ### Tech Stack: - Python ### Why It Matters: This project demonstrates the potential of combining CNNs and SVMs for image classification, showcasing the effectiveness of this hybrid approach. The results suggest that the CNN-SVM model could be a valuable addition to the toolkit for image classification tasks, potentially leading to improved accuracy and performance.

Overview: Context Engine is an AI coding assistant that leverages 30+ MCP tools for semantic search, symbol graph navigation, memory, and more. It provides guidance on how to use these tools effectively to enhance code efficiency and accuracy. Key features: - Semantic code search - Memory management - Symbol graph navigation - Multi-hop impact analysis - Cross-repo search capabilities Tech stack: Python Why it matters: Context Engine simplifies the process of finding, understanding, and managing code by providing a unified interface for various MCP tools, making it easier for developers to optimize their projects and improve their productivity. Repository: Context-Engine-AI/Context-Engine GitHub stars: 401 GitHub description: Context-Engine MCP - Agentic Context Compression Suite GitHub description: Context Engine is an AI coding assistant that leverages 30+ MCP tools for semantic search, symbol graph navigation, memory, and more. It provides guidance on how to use these tools effectively to enhance code efficiency and accuracy. Context Engine simplifies the process of finding, understanding, and managing code by providing a unified interface for various MCP tools, making it easier for developers to optimize their projects and improve their productivity.