Projects

### Overview: This repository is a curated list of datasets, codebases, and papers on Multi-Task Learning (MTL), focusing on a Machine Learning perspective. It provides an in-depth survey of the field, including benchmarks, datasets, codebases, architectures, optimization strategies, and theoretical aspects of multi-task learning. ### Key features: - Datasets: MultiMNIST, NYUv2, CityScapes, PASCAL Context - Codebases: Various implementations and examples - Architectures: Hard Parameter Sharing, Soft Parameter Sharing, Decoder-focused Model, Modulation & Adapters, Modularity, MoE, Routing & NAS, Task Representation, Task Relationship Learning: Grouping, Tree (Hierarchy) & Cascading, Theory, Misc ### Tech stack: Unknown ### Why it matters: This repository serves as a valuable resource for researchers and practitioners interested in multi-task learning, providing a comprehensive overview of the field and practical implementations. It is particularly useful for those looking to understand the theoretical foundations and practical applications of multi-task learning in various domains such as computer vision, natural language processing, and robotics.

### Overview: iNLTK is a Natural Language Toolkit designed specifically for Indic languages, offering developers out-of-the-box support for various NLP tasks. ### Key features: - Support for multiple Indic languages including Hindi, Punjabi, Gujarati, Kannada, Malayalam, Oriya, Marathi, Bengali, Tamil, Urdu, Nepali, Sanskrit, English, Telugu, and mixed scripts like Hinglish and Tanglish. - Comprehensive documentation and installation instructions available at https://inltk.readthedocs.io. - Emphasizes practical applications for application developers by providing ready-to-use models and tools for NLP tasks. ### Tech stack: Python ### Why it matters: iNLTK aims to simplify the development process for NLP tasks in Indic languages, making it easier for developers to integrate language-specific NLP capabilities into their applications.

Overview: OpenCOOD is an open-source cooperative detection framework designed for autonomous driving, officially supported by the ICRA 2022 paper "OPV2V". Key features: - Easy data API for multiple popular multi-agent perception datasets. - Support for LiDAR, Camera, and Radar APIs. - SOTA 3D detection backbones (PointPillar, Pixor, VoxelNet, SECOND). - Multiple sparse convolution versions (Spconv 1.2.1, Spconv 2.x). - Support for SOTA multi-agent perception models (Attentive Fusion, Cooper, F-Cooper, V2VNet, CoAlign, FPV-RCNN, DiscoNet, CoBEVT, Where2comm, V2VAM). Tech stack: Python Why it matters: OpenCOOD provides a comprehensive and efficient solution for cooperative detection in autonomous driving, making it a valuable resource for researchers and developers in the field.

### Overview: This project is a Graph RAG System specifically designed for medical information retrieval, focusing on the Medical-Graph-RAG repository. It aims to provide a robust and evidence-based approach to medical information retrieval. ### Key Features: - **Graph RAG System**: A specialized Graph RAG model tailored for the medical domain. - **GitHub Stars**: 839, indicating significant interest and adoption. - **Paper Link**: [ACL 2025](https://arxiv.org/abs/2408.04187), detailing the methodology and results. - **Demo**: A Docker-based demo available, allowing users to interact with the system through a web-based PubMed search. - **Quick Start (Baseline)**: A simple Graph RAG pipeline setup using a specific dataset and prompt. - **Build from Scratch**: Instructions for constructing the Graph RAG flow from scratch, including dataset preparation and model inference. ### Tech Stack: - Python - Neo4j - LLMs (ChatGPT, OpenAI API, etc.) ### Why It Matters: This project offers a practical solution for medical information retrieval, leveraging advanced graph-based models to enhance the accuracy and relevance of medical data retrieval. The use of Neo4j for graph construction and LLMs for prompt-based inference makes it highly efficient and scalable.

Overview: Yatai is a Kubernetes deployment operator designed for integrating BentoML into GitOps workflows, enabling efficient deployment and scaling of Machine Learning services on any Kubernetes cluster. Key features: - Seamless integration with BentoML - Kubernetes-native workflow via BentoDeployment CRD - Cloud native and DevOps friendly - Optimized for CI/CD and DevOps Tech stack: TypeScript Why it matters: Yatai empowers developers to deploy BentoML on Kubernetes, facilitating integration into CI/CD pipelines and streamlined DevOps workflows.

### Overview: This repository, **Awesome-LLM-Uncertainty-Reliability-Robustness**, is a curated list of resources and papers focusing on **Uncertainty**, **Reliability**, and **Robustness** in **Large Language Models**. It aims to provide a comprehensive overview of the current state of research in these areas, highlighting the challenges and solutions being explored. ### Key features: - **Uncertainty Estimation** - **Calibration** - **Ambiguity** - **Confidence** - **Hallucination** - **Mechanistic Interpretability** - **Truthfulness** - **Reasoning** - **Prompt tuning, optimization and design** - **Instruction and RLHF** - **Tools and external APIs** - **Fine-tuning** - **Invariance** - **Distribution Shift** - **Out-of-Distribution** - **Adaptation and Generalization** - **Adversarial** - **Attribution** - **Causality** ### Tech stack: Python, Markdown, Git, and various academic papers and resources. ### Why it matters: The development of large language models (LLMs) is crucial for the future of AI, but they face significant challenges in terms of reliability, uncertainty, and robustness. This repository aims to provide a platform for researchers, developers, and practitioners to share their insights, ideas, and solutions in these areas, ultimately contributing to the advancement of AI technology.

## Overview: Translate is a PyTorch-based library designed for training sequence-to-sequence models in machine translation. It leverages the fairseq library for general sequence-to-sequence tasks and supports exporting models to Caffe2 for production use. ## Key Features: - **Training of Sequence-to-Sequence Models**: Utilizes PyTorch for training sequence-to-sequence models. - **Fairseq Library Integration**: Built on fairseq, allowing seamless integration with fairseq models. - **ONNX Export Support**: Ability to export models to Caffe2 graphs via ONNX for production use. - **Separate Export of Components**: Export encoder and decoder components to Caffe2 separately. - **Beam Search Implementation**: Beam search is implemented in C++. - **Future Plans**: Plans to add support for exporting beam search as well. ## Tech Stack: PyTorch ## Why it Matters: Translate is a valuable tool for developers working on machine translation projects, offering a robust framework for training sequence-to-sequence models and the ability to export models for production use.

Overview: This project is a concise implementation of a Twitter/Cornell-Movie Chatbot using 200 lines of code and TensorLayer. Key features: - Implementing a chatbot using TensorLayer - Using 200 lines of code for simplicity - Incorporating references to key papers on sequence-to-sequence models Tech stack: Python Why it matters: This project serves as a practical example of how to build a chatbot using a specific model architecture and demonstrates the effectiveness of using TensorLayer for such tasks.

Overview: DiffuSeq is a powerful sequence-to-sequence language model that achieves state-of-the-art performance in text generation tasks. It is trained end-to-end using a conditional diffusion language model approach. Key features: - Conditional language model trained end-to-end - Establishes connections between AR, NAR, and DiffuSeq models - Accelerates training convergence by 4x and generates samples 800x faster - Matches or surpasses competitive AR, iterative NAR, and large-PLMs in quality and diversity Tech stack: Python Why it matters: DiffuSeq represents a significant advancement in sequence-to-sequence learning paradigms, offering improved performance and efficiency in text generation tasks.

Overview: 这套仓库是为小白用户准备的面向CTF/逆向技能的整合包，旨在帮助用户快速上手并解决第一步不知道做什么的问题。 Key features: - 自动判断题型 - 头脑风暴功能 - 三种风格的教学模式 - 可以切换到提示模式 - 路由到合适的技能模块 Tech stack: Python Why it matters: 为初学者提供了一个统一入口，帮助他们快速上手CTF/逆向技能，无需过多的术语和复杂的流程。 Repository: asdfgh1445/ctf-super-hub Primary language: Python GitHub stars: 838 GitHub description: 面向小白用户的CTF / 逆向Skills整合包：自动分流、头脑风暴、教学模式、比赛模式、只提示模式 README excerpt: <div align="center"> <p> <img src="./assets/banner.svg" alt="CTF Super Hub" width="100%" /> </p> # CTF Super Hub / CTF Skills 小白整合包 **给小白用户准备的一套CTF / 逆向题目统一入口。** 如果你经常遇到这些问题： **题目不会分类、skill不知道怎么选、第一步不知道该做什么。** 那这个仓库就是为你准备的。 <p> <a href="./START-HERE.md">快速开始</a> · <a href="./SKILL-INDEX.md">技能索引</a> · <

### Overview: This repository contains a Python implementation of several common Connectionist Temporal Classification (CTC) decoding algorithms, including best path, beam search, lexicon search, prefix search, and token passing. The minimalistic language model provided is designed to enhance decoding accuracy. ### Key Features: - Best Path Decoder: The fastest algorithm, suitable for most applications. - Beam Search Decoder: Integrates a character-level language model, allowing for improved reading accuracy. - Lexicon Search Decoder: Computes a first approximation using best path decoding, then retrieves similar words using a BK-tree. - Prefix Search Decoder: A basic approach for decoding. - Token Passing Algorithm: An alternative to beam search for certain applications. ### Tech Stack: Python ### Why It Matters: This implementation is useful for developers looking to implement CTC decoding algorithms in their projects, providing a minimalistic language model to enhance performance. The provided scripts in the `tests/` folder offer examples of how to use the decoders effectively. ### Repository: - **GitHub Stars:** 838 - **Repository:** githubharald/CTCDecoder - **Primary Language:** Python ### Installation: - **Steps:** 1. Go to the root level of the repository. 2. Execute `pip install .` to install the package. 3. Run `pytest` in the `tests/` folder to check installation. ### Usage: - **Basic Usage:** ```python import numpy as np from ctc_decoder import best_path, beam_search mat

### Overview: Attestix is an open-source project that provides AI agent identity, signed credentials, delegated authority, and portable governance evidence. It includes Python core and offline verifier SDKs for secure AI governance workflows. ### Key Features: - **Signed Identity and Evidence**: Records cryptographically verifiable evidence to support AI governance workflows. - **Open-Source Identity, Credentials, Compliance Automation, and Trust Scoring**: Provides a robust solution for AI agent identity and governance. - **MCP and REST Interfaces**: Supports MCP and REST interfaces for integration with various systems. - **Automated Functional Tests and Attestix Interoperability Benchmarks**: Ensures compatibility and interoperability with other tools. - **Real Integrations with LangChain, OpenAI Agents SDK, and CrewAI**: Enables seamless integration with popular AI tools. ### Tech Stack: Python ### Why it Matters: Attestix offers a secure and transparent way to manage AI agent identities and governance, ensuring compliance and accountability in AI-driven workflows. Its open-source nature and community-driven approach make it an ideal choice for developers looking to implement secure AI governance solutions.

### Overview: VoxPoser is a method that uses large language models and vision-language models to synthesize zero-shot trajectories for robotic manipulation tasks. This project provides a demo implementation of VoxPoser in the RLbench environment, aiming to showcase the capabilities of the method without requiring any training data. The code is designed to be easily deployable on real robots, with a focus on real-time mapping and object recognition. ### Key Features: - **Zero-Shot Trajectory Synthesis**: VoxPoser generates trajectories for robotic manipulation tasks without the need for any training data. - **Composable 3D Value Maps**: The method uses 3D value maps to decompose complex tasks into simpler sub-tasks, making it highly modular and adaptable. - **Real-Time Mapping**: The implementation uses object masks provided by RLbench, ensuring real-time performance. ### Tech Stack: - Python ### Why It Matters: VoxPoser offers a practical solution for robotic manipulation tasks by leveraging language models to synthesize zero-shot trajectories. This approach is particularly useful in scenarios where training data is scarce or expensive, making it a valuable tool for researchers and practitioners in robotics and AI. Repository: [huangwl18/VoxPoser](https://github.com/huangwl18/VoxPoser) GitHub Stars: 836

### Overview: OKF Agent Memory is a Git-native persistent project memory system designed for AI agents. It addresses the challenges of traditional approaches by providing a dual-memory architecture that reduces context bloat and ensures semantic search capabilities. ### Key features: - **Dual-Memory Agent Architecture (DMAA)**: Combines a normative working memory (push layer) and a semantic domain memory (pull layer) to ensure zero-tolerance invariants and efficient retrieval. - **OKF v0.2 Knowledge Bundle**: A knowledge bundle that consumes 0 tokens at baseline and is queried in microseconds, ensuring zero external dependencies. - **Zero-Tolerance Invariants**: Enforced by a permanent behavioral codex (capped at 400 tokens) in the normative working memory layer. - **Selective Retrieval**: Agents can selectively retrieve context and facts from the semantic domain memory layer, enhancing semantic search capabilities. ### Tech stack: Go 1.26 with zero external dependencies. ### Why it matters: By providing a compact, semantically enriched working memory and a persistent, zero-dependency knowledge bundle, OKF Agent Memory slashes token bloat by 80% and eliminates the need for external databases or dependencies, making it highly efficient and scalable for AI agents.

Overview: DeepInverse is a PyTorch-based library designed for solving imaging inverse problems using deep learning. It accelerates research across imaging domains and enhances reproducibility through a modular framework. Key features: - A large framework of predefined imaging operators - State-of-the-art deep neural networks for reconstruction and denoising - Comprehensive frameworks for iterative restoration, optimization, and unfolded architectures - Training losses for inverse problems - Sampling algorithms and diffusion models for uncertainty quantification - A framework for building datasets for inverse problems Tech stack: Python Why it matters: DeepInverse accelerates deep learning research across imaging domains, enhances research reproducibility via a common modular framework, and lowers the entrance bar to new practitioners. Repository: deepinv/deepinv GitHub stars: 834 GitHub description: DeepInverse: a PyTorch library for solving imaging inverse problems with deep learning.

Overview: LlamaFarm is an open-source AI platform that allows developers to deploy AI models, agents, databases, RAG, and pipelines locally or remotely, providing complete privacy and no API costs. Key features: - 🔒 Complete Privacy - 💰 No API Costs - 🌐 Offline Capable - ⚡ Hardware Optimized Tech stack: Python Why it matters: LlamaFarm empowers developers to harness enterprise-level AI capabilities on their own hardware, ensuring data privacy and eliminating cloud-based API costs.

### Overview: Build real-time speech-to-text web applications using OpenAI's Whisper, focusing on 99 languages with faster-whisper, Diart, and Pyannote. ### Key features: - Real-time speech-to-text functionality - Support for multiple languages - Utilizes faster-whisper, Diart, and Pyannote libraries - Compatible with various computing devices ### Tech stack: Python ### Why it matters: This repository provides a comprehensive solution for developers looking to build real-time speech-to-text applications, making it easier to integrate and deploy such solutions quickly and efficiently.

### Overview: Scenario for Blender is a Python extension designed to bring advanced features such as AI images, video, 3D, PBR materials, audio, and 360 environment generation directly into Blender. This extension allows users to explore scene captures and selected meshes, creating a greybox layout with Blockout, and access a compact composer, sidebar, Studio view, and local MCP server for agents. The project is licensed under the GPL-3.0-or-later and is compatible with Blender 5.0 and later versions. ### Key Features: - **AI Images and Video**: Generate high-quality AI-generated images and videos. - **3D and PBR Materials**: Create and apply PBR materials with advanced lighting effects. - **Audio Integration**: Incorporate audio elements into the scene. - **360 Environment**: Generate and integrate 360-degree environment views. - **Scene Description to Layout**: Convert written scene descriptions into a greybox layout using Blockout. - **Local MCP Server**: Access a local MCP server for agents within the scene. ### Tech Stack: - Python ### Why It Matters: Scenario for Blender offers a seamless integration of advanced features directly into Blender, enhancing the creative workflow by providing a compact and efficient environment for scene creation, exploration, and collaboration. The ability to generate high-quality AI images, 3D models, and PBR materials, along with the integration of audio and 360 environment views, makes it a valuable tool for artists and designers looking to streamline their creative processes.

Overview: Eesen is an open-source project aimed at simplifying the ASR pipeline by replacing complex models like HMMs and GMMs with a single RNN, making it more straightforward and efficient. Key features: - Bi-directional RNNs with LSTM units for acoustic modeling. - CTC training objective for training speed. - Principled decoding approach based on WFSTs. - RNN language model decoding for flexibility. Tech stack: C++ Why it matters: Eesen simplifies the ASR pipeline by eliminating unnecessary components, making it easier to implement and potentially faster to train, while still achieving high accuracy.

### Overview: Friday is a private control plane designed for AI agents, providing a seamless and secure way to delegate tasks to cloud-based AI models. It ensures that your goals are executed correctly and transparently, without the need for constant monitoring. ### Key Features: - **Local-first Kernel**: Runs your AI agents locally, ensuring your data and keys are always kept secure. - **BYOK (Bring Your Own Key)**: Allows you to use your own AI models without needing to pay for them. - **Approval-First**: Checks the work of AI agents before they perform risky tasks, ensuring safety and accountability. - **Human-Controlled**: You retain control over the AI agents, making decisions and verifying their work. - **Evidence-Backed**: Agents must provide proof of completion for tasks to be considered completed. ### Tech Stack: TypeScript ### Why It Matters: Friday provides a secure and efficient way to delegate tasks to AI agents, ensuring that your goals are executed correctly and transparently. It allows you to delegate tasks to cloud-based AI models without the need for constant monitoring, making it ideal for those looking to automate their workflows while maintaining control over the process. ### Repository: GitHub Repository: thesongzhu/Friday ### GitHub Stars: 836 ### Primary Language: TypeScript ### GitHub Description: Your private control plane for AI agents. Hand it a goal. It routes each step to an AI, checks the work, asks before anything risky, and remembers what you told it — so you're steering, not babysitting. You hold the keys. Local