Projects

### Overview: Giveme5W1H is an open-source system designed to extract the journalistic five W and one H questions (who, what, when, where, why, and how) from news articles. This tool simplifies the process of extracting key information from articles, making it accessible through a RESTful API or as a Python library. ### Key Features: - Extracts 5W1H phrases from news articles. - Supports access via RESTful API or as a Python library. - Requires the Stanford CoreNLP Server to be installed and running. - Provides sample Python scripts and a sample JSON file for extracting 5W1Hs from multiple articles. ### Tech Stack: Python 3.6+ with the Stanford CoreNLP Server. ### Why it Matters: Giveme5W1H simplifies the task of extracting key information from news articles, making it easier for developers and journalists to quickly understand the main events and details of a news story.

Overview: PHP Text Analysis is a comprehensive library designed for performing Information Retrieval (IR) and Natural Language Processing (NLP) tasks using the PHP programming language. Key features: - Document classification - Sentiment analysis - Compare documents - Frequency analysis - Tokenization - Stemming - Collocations with Pointwise Mutual Information - Lexical diversity - Corpus analysis - Text summarization Tech stack: PHP Why it matters: PHP Text Analysis provides a robust set of tools for text analysis and processing, making it an essential resource for developers looking to enhance their applications with natural language capabilities.

### Overview: This project implements Generative Adversarial Networks (GANs) for extreme learned image compression using TensorFlow. The method was developed by Agustsson et al. in their paper "Generative Adversarial Networks for Extreme Learned Image Compression" (arXiv:1804.02958). ### Key Features: - **Generative Adversarial Networks (GANs)**: The project leverages GANs for image compression. - **TensorFlow Implementation**: The code is written in Python using TensorFlow 1.8. - **Bottleneck Channels**: The model uses a channel bottleneck of 8 channels for compression, resulting in a bitrate of 0.072 bpp. - **Global Compression**: The method compresses images without semantic maps, achieving a compression rate of 0.072 bpp. - **Results**: The project includes results from experiments on the Cityscapes dataset, demonstrating the effectiveness of the compression technique. ### Tech Stack: Python ### Why it Matters: The proposed method offers a novel approach to image compression that significantly reduces the bitrate while maintaining perceptual quality. This technique could be particularly useful in applications where high compression is necessary while preserving the visual integrity of images. ### Repository: Repository: Justin-Tan/generative-compression GitHub Stars: 533 GitHub Description: TensorFlow Implementation of Generative Adversarial Networks for Extreme Learned Image Compression ### Primary Language: Python ### GitHub Description: The project is a TensorFlow implementation of Generative Adversarial Networks

Overview: Poke-env is a Python library designed for building scripted agents, self-play experiments, and reinforcement learning workflows on the Pokemon Showdown platform. Key features: - Self-play experiments - Reinforcement learning workflows - Access to a Pokemon Showdown server for training and development Tech stack: Python Why it matters: Poke-env simplifies the process of creating and testing bots on the Pokemon Showdown platform, making it easier to develop and experiment with AI agents in a competitive environment. Repository: hsahovic/poke-env GitHub stars: 532 GitHub description: Poke-env: Python Interface for Pokemon Showdown Bots

Overview: This repository contains implementations of various NLP tasks, including new word discovery, word embeddings, text classification using PyTorch, named entity recognition (NER), text summarization, sentence similarity judgment, triple extraction, pre-trained models, and more. Key features: - New Word Discovery - Word2Vec (Skip-gram) - GloVe - TextCNN - FastText - TextRCNN - TextRNN_Att - DPCNN - XGBoost - Distill_finetune_Bert - Pattern-Exploiting-Training - R-Drop - Bert-MRC - Bert-CRF - Bert-Label-Semantics - Bert-MLM - Seq2seq - Seq2seq_Att - Transformer - GPT - Bert-seq2seq - Bert-extractive-summarizer - Sentence-Similarity - MultiLabel-Classification - Relation-Extraction - Pretrained-Language-Model - P-tuning V1 - QA Tech stack: Python Why it matters: This repository provides a comprehensive set of tools and models for various NLP tasks, making it a valuable resource for developers looking to improve their NLP skills and projects.

### Overview: `language_tool_python` is a Python wrapper for LanguageTool, a free, non-AI, open-source grammar, style, and spell checker. This library allows developers to detect and correct errors in their Python scripts or command-line inputs, using a local Java server or the public LanguageTool API. ### Key Features: - Detect and fix grammar, style, and spelling errors in Python scripts. - Supports multiple languages and locales. - Integrates seamlessly with Python scripts and command-line tools. - Easy to install and use via pip. ### Tech Stack: Python ### Why It Matters: For developers who need to ensure the quality of their Python code, `language_tool_python` provides a reliable and efficient way to catch and correct errors. It's particularly useful for projects that require a high level of code review or automated checks, ensuring that the codebase remains clean and maintainable. ### Repository: jxmorris12/language_tool_python - GitHub Stars: 533 - Primary Language: Python - GitHub Description: a free, non-AI python grammar checker 📝✅

Overview: This repository is a curated list of AI-powered browser extensions, plugins, and add-ons designed to enhance productivity and efficiency in various tasks such as writing, research, automation, translation, coding, and browsing. Key features: - AI-driven browser automation and workflow tools - AI writing and editing assistants - AI-powered research and summarization tools - AI translation capabilities - AI-enhanced email and communication tools - AI-assisted productivity and meeting tools Tech stack: JavaScript, TypeScript, and Webpack Why it matters: These extensions and tools leverage AI to streamline tasks, improve accuracy, and enhance overall productivity, making them valuable additions to any developer's toolkit. Repository: kklt92/awesome-ai-extensions GitHub stars: 533 GitHub description: A curated list of AI-powered browser extensions, plugins, and add-ons.

Overview: 首个中医大语言模型“仲景”，由张仲景智慧启发，旨在为传统中医领域提供专业且可靠的工具。 Key features: - 多任务诊疗分解策略 - 13.5万+专业指令数据 - 跨专科泛化能力 - 人类医师评估验证 - 轻量级部署友好 Tech stack: Jupyter Notebook Why it matters: 该模型通过借鉴人类记忆与学习机制，构建了高质量的指令数据，旨在传承古代中医智慧与现代技术创新，为医学领域提供专业且可靠的工具。

Overview: VectorHub is a free and open-source learning hub for developers interested in vector retrieval, offering practical resources to create MVPs, solve use case-specific challenges, and make vector retrieval practical and useful. Key features: - Practical learning materials for creating MVPs - Resources to solve specific challenges in vector retrieval - Guidance on moving MVPs to production and making them useful - Information about vendors in the vector retrieval space Tech stack: Jupyter Notebook Why it matters: VectorHub provides a comprehensive and practical resource for developers looking to integrate vector retrieval into their ML stack, offering a step-by-step approach to creating, solving problems, and making vector retrieval useful. Repository: superlinked/VectorHub GitHub stars: 533 GitHub description: Deprecated historical repo. Superlinked now develops SIE, a self-hosted inference engine for embeddings, reranking, OCR, extraction, and document processing.

Overview: This project presents a neural method for reconstructing 3D scenes using the Manhattan-world assumption, focusing on the CVPR 2022 Oral presentation. It includes detailed instructions for running on custom data and provides quantitative and qualitative results for comparison with other models. Key features: - Neural 3D Scene Reconstruction - Manhattan-world Assumption - Custom Data Instructions - ScanNet Scene Data Evaluation - Mesh Extraction and Evaluation Tech stack: Python Why it matters: This project offers a novel approach to 3D scene reconstruction that leverages the Manhattan-world assumption, making it particularly useful for applications requiring efficient and accurate reconstruction. Repository: zju3dv/manhattan_sdf Primary language: Python GitHub stars: 533 GitHub description: Code for "Neural 3D Scene Reconstruction with the Manhattan-world Assumption" CVPR 2022 Oral

Overview: Explore a comprehensive collection of resources, tutorials, papers, tools, and best practices for fine-tuning Large Language Models (LLMs) like GPT, BERT, RoBERTa, and their numerous variants. Perfect for ML practitioners and researchers! Key features: - Tutorials - Papers - Tools - Frameworks - Best practices Tech stack: Python, JavaScript, Ruby, and various machine learning libraries. Why it matters: Fine-tuning LLMs is crucial for enhancing their performance and making them suitable for specific applications, ensuring alignment with the task's requirements and context.

### Overview This project implements the Single Shot MultiBox Detector (SSD) in TensorFlow for traffic sign detection and classification. The model has been trained on the LISA Traffic Sign Dataset, focusing on detecting and classifying stop signs and pedestrian crossing signs. The implementation aims to achieve high detection rates and real-time performance, although it currently struggles with model overfitting. ### Key Features - Traffic sign detection and classification - Real-time performance on a GTX 1080 with an Intel Core i7-6700K - Currently trained on stop signs and pedestrian crossing signs - Supports inference on images and videos ### Tech Stack - Python 3.5+ - TensorFlow v0.12.0 - Pickle - OpenCV-Python - Matplotlib (optional) ### Why It Matters This project is a practical implementation of SSD for traffic sign detection, showcasing how to leverage TensorFlow for real-world applications. It demonstrates the model's ability to perform well on a specific dataset and provides insights into improving performance through data augmentation and pre-training on larger datasets. ### Repository Repository: georgesung/ssd_tensorflow_traffic_sign_detection GitHub Stars: 532 GitHub Description: Implementation of Single Shot MultiBox Detector in TensorFlow, to detect and classify traffic signs. This implementation was able to achieve 40-45 fps on a GTX 1080 with an Intel Core i7-6700K. ### Dependencies - Python 3.5+ - TensorFlow v0.12.0

Overview: TorchUncertainty is an open-source framework designed to enhance the reliability of deep learning models through uncertainty quantification techniques. It aims to provide a comprehensive set of methods and tools for researchers and developers to improve the performance and reliability of their models. Key features: - Uncertainty-aware training and evaluation routines for classification, regression, segmentation, and pixelwise regression. - Automated evaluation of models with proper scores, calibration, selective classification, out-of-distribution detection, and distribution-shift metrics. - Layers, models, metrics, and losses ready to be integrated into existing networks. - Scikit-learn-style post-processing methods for temperature scaling and conformal prediction. - Transformations and augmentations, including a set of corruptions for building "corrupted datasets" on HuggingFace. Tech stack: Python Why it matters: TorchUncertainty is essential for researchers and developers looking to improve the reliability and performance of their deep learning models. It provides a robust set of tools and methods to evaluate and enhance the uncertainty of models, making them more robust and reliable in various applications. Repository: torch-uncertainty/torch-uncertainty GitHub stars: 532 GitHub description: Open-source framework for uncertainty and deep learning models in PyTorch 🌱

### Overview: ConvMAE: Masked Convolution Meets Masked Autoencoders is a self-supervised framework that demonstrates the effectiveness of multi-scale hybrid convolution-transformer for discriminative representation learning. The project is based on the NeurIPS 2022 paper "MCMAE: Masked Convolution Meets Masked Autoencoders," which showcases ConvMAE's superior performance across various downstream tasks such as ImageNet pretraining, object detection, and semantic segmentation. The project includes pre-trained checkpoints for ImageNet-1K pretraining, Object Detection, and Semantic Segmentation, and supports fast pretraining with significantly reduced computational requirements. ### Key Features: - **ImageNet Pretrain**: Pretrained checkpoints and logs for ImageNet-1K pretraining. - **ImageNet Finetune**: Pretrained models and logs for ImageNet-1K finetuning. - **Object Detection**: Object Detection with Mask-RCNN. - **Semantic Segmentation**: Semantic Segmentation with UperNet. - **Video Classification**: Supports video classification tasks. ### Tech Stack: - Python ### Why It Matters: ConvMAE demonstrates that multi-scale hybrid convolution-transformer can learn more discriminative representations via the mask auto-encoding scheme, showcasing strong and efficient self-supervised learning capabilities. The project's pre-trained checkpoints and logs provide a robust foundation for various downstream tasks, making it a valuable resource for researchers and practitioners in computer vision and deep learning.

### Overview: MedCAT is a medical concept annotation tool designed to extract information from Electronic Health Records (EHRs) and link it to biomedical ontologies such as SNOMED-CT and UMLS. It is part of the Cogstack ecosystem and can be used to train and download models for various biomedical ontologies. ### Key features: - **Model Packs**: Available models include UMLS Small, SNOMED International, UMLS Dutch v1.10, and UMLS Full. - **Model Training**: Models are trained on MIMIC-III datasets. - **Documentation**: Official documentation and a discussion forum are available. - **Public Models**: Four public models are available: UMLS Small, SNOMED International, UMLS Dutch v1.10, and UMLS Full. ### Tech stack: Python ### Why it matters: MedCAT is a valuable tool for medical researchers and practitioners who need to extract meaningful information from EHRs and link it to biomedical ontologies. It is part of the Cogstack ecosystem, which provides a comprehensive set of tools and resources for medical research and development.

Overview: This project focuses on the implementation and research of Kolmogorov-Arnold convolutional networks, including 1D, 2D, and 3D convolutions, ResNet-like and DenseNet-like models, and training code using PyTorch with CUDA acceleration. It evaluates performance on datasets like MNIST, CIFAR, Tiny ImageNet, and Imagenet1k. Key features: - Implementations of 1D, 2D, and 3D convolutions with different kernels - ResNet-like and DenseNet-like models - Training code based on PyTorch with CUDA acceleration - Scripts for experiments with CIFAR-10 and Tiny ImageNet Tech stack: Python Why it matters: This project provides a comprehensive framework for training, validating, and quantizing Kolmogorov-Arnold convolutional networks, making it valuable for researchers and developers looking to explore this innovative model architecture. Repository: IvanDrokin/torch-conv-kan Primary language: Python GitHub stars: 531

### Overview: EdgeYOLO是一个在嵌入式设备上实现的高性能无锚框目标检测器，能够在NVIDIA Jetson AGX Xavier上达到34FPS，并在COCO2017和VisDrone2019-DET数据集上分别获得50.6%和25.9%的AP准确度。 ### Key features: - 边缘设备友好的无锚框目标检测器 - 较小模型EdgeYOLO-S在COCO2017数据集上达到了53FPS的速度 - 支持SAM辅助的图像标注工具SAMLabeler Pro - 更强大的数据增强方法 - 使用RH损失函数提升中小模型的检测效果 - 论文预印版已在arxiv公布 ### Tech stack: Python ### Why it matters: EdgeYOLO在嵌入式设备上实现了高性能的无锚框目标检测，适用于各种边缘计算场景，具有广泛的应用潜力。

Overview: This project is a speech recognition library for the browser, built using a WebAssembly build of Vosk, making it easy to integrate speech recognition functionality into web applications. Key features: - Real-time speech recognition - Cross-browser compatibility - Supports multiple languages - Uses WebAssembly for efficient performance Tech stack: JavaScript, WebAssembly Why it matters: This library simplifies the integration of speech recognition into web applications, making it a valuable tool for developers looking to add speech recognition capabilities to their projects without the complexity of native development.

### Overview: Talk2Arxiv is a tool that allows users to read and interact with any paper on arXiv or bioRxiv by transforming the paper into a clean HTML format. It supports both HTML and PDF versions, providing a seamless reading experience with highlighting and chat functionality. ### Key features: - **Clean HTML Reader:** Papers load from arXiv's HTML5 version (with ar5iv as a fallback) or bioRxiv's full-text pages, including a table of contents and reading progress. - **PDF Fallback:** Papers without HTML versions open as PDFs rendered with EmbedPDF (PDFium in WebAssembly), with highlighting and chat capabilities. - **Highlight and Ask:** Select any passage and click "Ask AI" or "Explain" to send the passage to the chat with LaTeX equations. - **Citation Previews:** Click a citation in an HTML paper to preview it in a card, or a bottom sheet on mobile. Adding a cited paper to chat. - **Mobile Reading:** A section tracker, glass composer, and chat sheet keep the paper in place, with drafts surviving chat sessions. - **Starting Prompts:** Summarize the paper, map its claims to evidence, or ask about method and limitations. - **Whole-Paper Context:** The full paper text goes into the model's context, with OpenAI's prompt cache reusing it for each question. - **Cached Answers:** Identical questions with the same paper and conversation reuse completed answers, including starting prompts. Partial, stopped, and failed generations are never cached. - **

### Overview: Watch Skill is a powerful tool that turns video, audio, and screen activity into searchable, timestamped evidence. It verifies work with deterministic contracts, ensuring results are verifiable and not based on model opinion. This agent workspace, built on DeepSeek Harness, is designed to be integrated into existing AI workflows. ### Key features: - **Video and Audio Analysis**: Converts video and audio content into searchable, timestamped evidence. - **Screen Activity Capture**: Captures and indexes screen activity with absolute timestamps. - **Contract Verification**: Verifies work with deterministic contracts, providing a clear, verifiable outcome. ### Tech stack: Python + npm + MCP + CLI + REST + Web ### Why it matters: Watch Skill provides a concrete, verifiable way to prove work, ensuring that the results are not based on subjective models but on concrete evidence. This makes it invaluable in professional environments where accountability and transparency are crucial. ### Repository: oxbshw/watch-skill ### GitHub stars: 530 ### GitHub description: Give AI agents eyes, ears, and verifiable results.