Projects

### Overview: This repository is a benchmark framework designed to evaluate and compare various speech-to-text engines. ### Key features: - **Data**: Provides datasets for benchmarking, including LibriSpeech, TED-LIUM, Common Voice, Multilingual LibriSpeech, VoxPopuli, Flours, JSUT-BASIC, and Zeroth Korean. - **Metrics**: Offers metrics such as Word Error Rate (WER), Character Error Rate (CER), Punctuation Error Rate (PER), Core-Hour, Word Emission Latency, and Model Size. - **Engines**: Includes popular engines like Amazon Transcribe, Azure Speech-to-Text, Google Speech-to-Text, IBM Watson Speech-to-Text, Moonshine, Nemotron 3.5 ASR Streaming, OpenAI Whisper, Vosk, Whisper.cpp, Picovoice Cheetah, and Picovoice Leopard. ### Tech stack: Python ### Why it matters: This benchmark framework helps developers and researchers evaluate the performance and efficiency of speech-to-text engines, ensuring that they can make informed decisions about which engine to use for their projects.

### Overview: EfficientSAM3 is a progressive hierarchical knowledge distillation method that compresses SAM3 into lightweight, edge-friendly models. It achieves fast promptable concept segmentation and tracking by fine-tuning on 5% SA1B data and SACap labels. ### Key Features: - **Lightweight Vision and Text Encoders**: Compresses SAM3's vision and text encoders into lightweight student models. - **Competitive Performance**: Maintains competitive performance on downstream benchmarks. - **Fast Promptable Segmentation and Tracking**: Achieves fast promptable concept segmentation and tracking. ### Tech Stack: Progressive Knowledge Distillation ### Why it Matters: EfficientSAM3 enables efficient and lightweight models for fast concept segmentation and tracking, making it suitable for edge devices and real-time applications.

### Overview: This repository contains the source code for LaneATT, a novel state-of-the-art lane detection model presented in the paper "Keep your Eyes on the Lane: Real-time Attention-guided Lane Detection" (CVPR 2021). The paper was accepted to CVPR 2021, and the code is available on GitHub under the repository name "lucastabelini/LaneATT". ### Key features: - Real-time Attention-guided Lane Detection - State-of-the-art performance on various datasets - Python implementation using PyTorch and Torchvision - Supports multiple backbones and datasets ### Tech stack: - Python 3.8 - PyTorch 1.6 - CUDA 10.2 ### Why it matters: This project is significant for researchers and developers interested in real-time lane detection, as it offers a high-performance solution that can be easily integrated into various applications, including autonomous vehicles and traffic monitoring systems.

### Overview: SkillSpec simplifies the process of making skills testable and verifiable, ensuring agents follow their instructions and prove their work. ### Key Features: - **Risk Reporting**: Run a command to get a risk report on agent follow-through, highlighting potential issues. - **Contract Creation**: Convert skills into structured contracts that the agent must follow, providing a record of their actions and progress. - **Doctor Integration**: Use Doctor to test skills and ensure they meet the required criteria. ### Tech Stack: Rust ### Why It Matters: SkillSpec provides a structured approach to making skills testable and verifiable, ensuring agents follow their instructions and prove their work. This helps in avoiding buried rules, growing prose, and missing steps, ultimately improving the reliability and trustworthiness of skills in AI-driven systems.

### Overview: ML-ProjectKart showcases a collection of 234+ projects across various domains of machine learning, deep learning, computer vision, natural language processing, and generative AI. This repository aims to provide a beginner-friendly and efficient platform to learn and master these algorithms through hands-on projects. ### Key Features: - A wide range of projects covering machine learning, deep learning, computer vision, natural language processing, and generative AI. - Projects are designed to be beginner-friendly and easy to follow. - Projects are available as Jupyter Notebooks, making them accessible and easy to understand. - The repository includes a variety of projects, ensuring a comprehensive learning experience. ### Tech Stack: The projects are built using Python, TensorFlow, PyTorch, and other popular machine learning libraries. ### Why it Matters: This repository serves as a valuable resource for anyone interested in machine learning, deep learning, computer vision, natural language processing, and generative AI. It provides a platform to learn and practice these technologies through hands-on projects, making it an excellent resource for both beginners and experienced developers.

### Overview: This repository contains the official code for the paper "Long-form factuality in large language models" by DeepMind. It includes datasets, evaluation tools, and a benchmarking pipeline for measuring the accuracy of models in generating long-form factual responses. ### Key features: - **LongFact**: A dataset of 2,280 fact-seeking prompts requiring long-form responses. - **Search-Augmented Factuality Evaluator (SAFE)**: An automatic evaluation tool for model responses in long-form factuality settings. - **F1@K**: A metric that extends the F1 score to long-form settings by using recall from human-preferred length. - **Experimentation pipeline**: A set of pipelines for benchmarking models like OpenAI and Anthropic using the LongFact dataset and SAFE evaluation tool. ### Tech stack: Python ### Why it matters: This repository provides a comprehensive solution for benchmarking and evaluating the long-form factuality capabilities of large language models, contributing to the field of language understanding and AI research.

Overview: This project is an AI-driven toolkit designed to streamline and automate the cross-border e-commerce operations process, making it more efficient and effective for both new and experienced professionals. Key features: - **Market Research Agent**: Analyzes market trends and competitive positioning. - **Product Strategy Agent**: Determines product positioning and pricing strategies. - **Listing Creative Agent**: Generates creative product listings with optimized keywords, titles, and descriptions. - **Product Image Agent**: Creates high-quality product images with pixel-level compliance checks. - **Product Video Agent**: Generates engaging product videos with AI scripts and voiceovers. - **Advertising Optimization Agent**: Optimizes ad campaigns by diagnosing four types of funnel and calculating profitability. - **Review Analysis Agent**: Analyzes customer reviews to identify attribution and improve product descriptions. Tech stack: Python Why it matters: By breaking down complex operations into reusable AI Agent workflows, this toolkit helps businesses achieve consistent, high-quality results without the need for manual intervention. It ensures that even new employees can produce professional-looking product content, improving overall efficiency and reducing errors. Repository: heqiangdiao-wq/AI-Cross-Border-Ecommerce-Agent-Toolkit Primary language: Python GitHub stars: 667 GitHub description: This repository contains a toolkit designed to streamline and automate cross-border e-commerce operations, making it more efficient and effective for both new and experienced professionals.

Overview: ComfyUI-IF_AI_tools is a set of custom nodes for ComfyUI that allows you to generate prompts using a local Large Language Model (LLM) via Ollama. This tool enables you to enhance your image generation workflow by leveraging the power of language models. Key features: - nanoGraphRAG - OCR-RAG ColPali & ColQwen (Bialdy) - Supervision Object Detection Florence2 - Endpoints xAI, Transformers - IF_Assistants System Prompts with Reasoning/Reflection/Reward Templates and custom presets Tech stack: Python Why it matters: This tool provides a powerful way to integrate language models into the image generation process, making it easier to create more detailed and contextually relevant prompts.

### Overview: SwarmClaw is an open-source, self-hosted AI agent runtime and multi-agent framework designed for autonomous AI agent swarms. It includes features such as durable agent memory, MCP tools, skills, delegation, schedules, and support for 23+ LLM providers, including Claude, GPT, Gemini, OpenRouter, and Ollama. ### Key Features: - **Durable Agent Memory:** Ensures agent data persists across sessions. - **MCP Tools:** Provides a comprehensive set of tools for managing and orchestrating agents. - **Skills and Delegation:** Enables agents to perform specific tasks and delegate responsibilities. - **Schedules:** Allows agents to execute tasks at specific times. - **23+ LLM Providers:** Supports a wide range of language models for seamless integration. ### Tech Stack: - TypeScript ### Why It Matters: SwarmClaw offers a practical alternative to Claude Code and LangChain, providing a robust and scalable solution for self-hosted multi-agent AI workflows. Its features make it ideal for developers looking to create and manage autonomous AI agent swarms. ### Repository: - Repository: `swarmclawai/swarmclaw` - GitHub Stars: 697 - GitHub Description: Open-source agent swarms with durable agent memory, MCP tools, skills, delegation, schedules, and 23+ LLM providers — a practical Claude Code and LangChain alternative. ### Screenshots: - **Org Chart:** Visualizes agent teams, delegation, and live activity. - **Agent Chat:** Shows a conversation with a

Overview: This repository contains the code and data interface for MMEB-V3, a comprehensive benchmark for evaluating omni-modality embedding models across text, image, video, audio, visual document, and agent-centric retrieval scenarios. MMEB-V3 extends MMEB-V2 toward a fuller modality setting by adding 111 new tasks, resulting in 190 tasks in total. It introduces three major new evaluation categories: Audio Tasks, Text Retrieval, and Agent Tasks. Additionally, MMEB-V3 introduces OmniSET, a diagnostic component for controlled analysis of modality effects and instruction-conditioned cross-modal retrieval behavior. Key features: - MMEB-V3: Comprehensive benchmark for evaluating omni-modality embedding models. - 190 tasks in total, including 111 new tasks. - Three major new evaluation categories: Audio Tasks, Text Retrieval, and Agent Tasks. - OmniSET: Diagnostic component for controlled analysis of modality effects and instruction-conditioned cross-modal retrieval behavior. Tech stack: Python Why it matters: MMEB-V3 aims to diagnose whether current embedding models can reliably follow modality-specific instructions, such as retrieving an image, video, audio clip, document, tool, GUI element, or memory item under explicit task constraints. This benchmark is crucial for evaluating the performance of embedding models across various modalities and scenarios. Repository: TIGER-AI-Lab/VLM2Vec GitHub stars: 696 GitHub description: This repo contains the code and data interface for MMEB-V

Overview: Whatlanggo is a natural language detection library written in Go, supporting 84 languages and providing fast, reliable, and script-aware language recognition. Key features: - Supports 84 languages - Written in Go - No external dependencies - Fast - Recognizes both language and script Tech stack: Go Why it matters: Whatlanggo offers a fast, reliable, and script-aware natural language detection solution for Go developers, making it an essential tool for applications requiring language identification.

Overview: This repository is dedicated to collecting data sets, papers, and open-source implementations in the field of Natural Language Processing (NLP), with a focus on sentiment analysis, emotion attribution, evaluation of objects and evaluative words, and evaluation of posts. Key features: - Papers and Codes - Deep Learning for Sentiment Analysis: A Survey - Ensemble Approach for Emotion Cause Detection with Event Extraction and Multi-Kernel SVMs - Event-Driven Emotion Cause Extraction with Corpus Construction - A Question Answering Approach to Emotion Cause Extraction - Bootstrap Method for Automatic Rule Acquisition on Emotion Cause Extraction - Emotion Cause Detection for Chinese Micro-Blogs Based on ECOCC Model - Emotion Cause Detection with Linguistic Constructions - Emotion Cause Extraction, A Challenging Task with Corpus Construction - Extracting Causes of Emotions from Text - Based on E-CNN neural network's emotion cause identification method - Based on sequence labeling model's emotion cause identification method - Based on text's emotion automatic attribution method research - A Co-Attention Neural Network Model for Emotion Cause Analysis with Emotional Context Awareness - Emotion-Cause Pair Extraction: A New Task to Emotion Analysis in Texts - Fine-grained Opinion Mining with Recurrent Neural Networks and Word Embeddings - Character-based BiLSTM-CRF Incorporating POS and Dictionaries for Chinese Opinion Target Extraction - A Unified Model for Opinion Target Extraction and Target Sentiment Prediction - Evaluation Object and Evaluation Word Extraction Research Summary - Using Deep Long Short-Term Memory

Overview: This repository contains pre-trained models of popular CNN architectures trained on the CIFAR-10 dataset. The models have been modified to work with the CIFAR-10 format and include weights for easy loading. Key features: - Pre-trained models of vgg11_bn, vgg13_bn, vgg16_bn, vgg19_bn, resnet18, resnet34, resnet50, densenet121, densenet161, densenet169, mobilenet_v2, and googlenet. - Weights of the models are shared, allowing users to load and use them directly. - Highly re-producible and readable code using PyTorch-Lightning. Tech stack: Python Why it matters: This repository provides a convenient way to access pre-trained models for CIFAR-10 classification tasks, making it easier to start projects without having to retrain models from scratch. The weights are shared, ensuring compatibility and ease of use across different projects.

Overview: Swin2SR is a state-of-the-art super-resolution model that leverages transformers for image restoration and compression artifacts reduction. It achieves remarkable performance across classical SR, color JPEG artifact reduction, and compressed input super-resolution tasks. Key features: - State-of-the-art performance in classical, lightweight, and real-world image super-resolution - Effective reduction of color JPEG compression artifacts - Top solution for compressed input super-resolution in the AIM 2022 Challenge Tech stack: Python Why it matters: Swin2SR represents a significant improvement over SwinIR, offering enhanced performance and practical applications in various image processing tasks, making it a valuable tool for researchers and developers in the field of computer vision and image restoration.

### Overview: Xyne is an AI-first Search & Answer Engine designed for work, offering an open-source alternative to Glean, Gemini, and MS Copilot. It connects to your applications, securely indexes your data, and provides a graph-based solution to find information and get up-to-date answers with sources. ### Key Features: - **Self-hosted**: Deploy anywhere - laptop, on-prem, cloud. - **Model Agnostic**: Plugs into any LLM of your choice via any cloud provider. - **Private & secure**: No training on your data or prompts, no telemetry. - **Permissions-aware**: Enforces your apps' existing permissions. - **High performance**: Multi-threaded data ingestion. ### Tech Stack: TypeScript ### Why It Matters: Xyne is the most secure and future-proof solution for bringing AI into your work, providing a private and responsible way to integrate AI models with your existing applications.

### Overview: SAM3DBody-cpp is a real-time 3D full-body reconstruction engine that leverages a single camera to generate per-person MHR body pose parameters, camera translation, and optionally full 3D mesh vertices and 70 body/hand keypoints. This project is built entirely in C++ and utilizes ONNX Runtime and ggml for inference, making it a pure C++ runtime solution. ### Key features: - **Real-time 3D Reconstruction**: SAM3DBody-cpp reconstructs the full-body pose of multiple people from a single camera feed. - **Pure C++ Runtime**: The engine is written entirely in C++, ensuring high performance and low overhead. - **Multi-person BVH Output**: Supports exporting multi-person motion-capture data in the standard BVH format. - **MPEG Avatar Export**: Generates MPEG Avatar Representation Format (.arfz) containers for each detected person, including facial blendshapes. - **CUDA GPU Support**: Pre-built ONNX models are available for CUDA GPUs, providing faster inference times. - **CPU-only Support**: Models are also available for CPUs, allowing for offline processing without GPU dependencies. ### Tech stack: C++ ### Why it matters: SAM3DBody-cpp is essential for developers and researchers looking to integrate real-time 3D reconstruction into their projects without the need for Python or complex dependencies. Its performance and ease of use make it suitable for both real-time applications and offline processing, making it a valuable tool in the field of computer vision and motion capture. Repository: AmmarkoV/SAM3

### Overview: A curated map of memory for AI agents, featuring systems, benchmarks, and research papers that enhance long-term context, persistent recall, and reasoning capabilities in large language models and multimodal agents. ### Key Features: - **Products**: Resources for building AI agents with memory layers. - **Benchmarks**: Evaluations of memory systems. - **Tutorials**: Orientations and guides for understanding memory in AI. - **Surveys**: Research on memory architectures. - **Papers - Nonparametric Memory**: Papers focusing on non-parametric memory systems. - **Papers - Parametric Memory**: Papers exploring parametric memory systems. - **Memory Security & Defense**: Techniques to protect agent memory from poisoning and abuse. ### Tech Stack: Python ### Why It Matters: This repository serves as a comprehensive resource for developers and researchers interested in enhancing the memory capabilities of AI agents, making it easier to evaluate and implement memory systems in their projects.

### Overview: This Python package provides easy-to-use implementations of various multi-object tracking algorithms, including CentroidTracker, IOUTracker, CentroidKF_Tracker, and SORT. It also supports OpenCV-based object detectors such as YOLOv3 and TensorFlow MobileNet SSD. ### Key Features: - CentroidTracker - IOUTracker - CentroidKF_Tracker - SORT ### Tech Stack: Python ### Why It Matters: This package simplifies the process of implementing and using multi-object tracking algorithms, making it accessible to developers and researchers alike.

### Overview: This repository introduces a unified sequence parallel approach called LongContextAttention, which combines the strengths of DeepSpeed-Ulysses-Attention and Ring-Attention to enhance the performance and versatility of long context transformers. ### Key Features: - **Unified Sequence Parallelism (USP)**: A hybrid approach that leverages the benefits of both DeepSpeed-Ulysses-Attention and Ring-Attention. - **FlashAttention Compatibility**: Supports FlashAttention versions 2.6.x and 2.7.x, including both v3 and v2 versions. It also supports running without FlashAttention, suitable for Non-Uniform Memory Access (NPU) hardware. - **Installation**: Supports installation with FlashAttention 2.6.x and 2.7.x, including v3 and v2 versions, and provides an option for hardware that does not support FlashAttention, using torch for computation. ### Tech Stack: - Python ### Why It Matters: - **Enhanced Performance**: Combines the strengths of DeepSpeed-Ulysses-Attention and Ring-Attention to achieve better performance and versatility. - **Versatility**: Provides a unified solution that can be applied to various hardware environments, including NPU, Intel XPU, and others. - **Compatibility**: Supports FlashAttention 2.6.x and 2.7.x, ensuring compatibility with different hardware and environments. ### GitHub Stars: 696 ### Primary Language: Python ### Repository: feifeibear/long-context-attention

### Overview: This repository is a curated list of research papers, code, and datasets related to 3D object detection using LiDAR-based methods. It covers various conferences and workshops, including CVPR, ICCV, ECCV, and others, as well as datasets like KITTI, nuScenes, Lyft, Waymo Open Dataset, and more. ### Key features: - Papers, code, and datasets related to 3D object detection using LiDAR-based methods. - Curated list of research papers from top conferences and workshops. - Comprehensive datasets for training, validation, and testing, including KITTI, nuScenes, Lyft, and Waymo Open Dataset. - Papers on end-to-end multi-view fusion, fully convolutional networks, voxel-based methods, and real-time 3D detection using 3D LiDAR. - Papers on object detection and classification in occupancy grid maps, real-time 3D detection for autonomous driving, and 3D object detection from LiDAR information. ### Tech stack: - Python - PyTorch - TensorFlow - OpenCV - NumPy - Other relevant libraries and frameworks. ### Why it matters: This repository provides a valuable resource for researchers, developers, and enthusiasts interested in advancing the field of 3D object detection using LiDAR-based methods. It offers a comprehensive overview of the latest research, datasets, and tools, making it an essential resource for anyone looking to stay up-to-date with the latest developments in this rapidly evolving field.