Projects

### Overview: NOVA is an autoregressive video generation model that achieves high efficiency without using vector quantization. It reformulates the video generation problem as non-quantized autoregressive modeling for both temporal and spatial predictions. ### Key Features: - **Non-Quantized Autoregressive Generation**: NOVA generates videos without vector quantization, offering high efficiency. - **Unified Modeling**: The model can perform both text-to-image and image-to-video tasks, showcasing its versatility. - **State-of-the-Art Performance**: NOVA achieves state-of-the-art results in both text-to-image and image-to-video tasks. ### Tech Stack: Python ### Why It Matters: NOVA provides a novel approach to video generation that is both efficient and versatile, making it a significant advancement in the field of image and video synthesis.

### Overview: The Ego4d dataset repository is a comprehensive collection of large-scale multi-modal video data, including 3D video recordings, spanning over 3700 hours of first-person video data. It serves as a benchmark for video learning and perception tasks, providing essential tools and resources for researchers and developers in the field. ### Key Features: - **Ego-Exo4D Dataset**: A large-scale multi-modal multi-view video dataset with time-synchronized videos of participants recorded with egocentric (first-person) and exocentric (third-person) perspective cameras. - **Ego4D Dataset**: The world's largest egocentric video ML dataset, featuring over 3700 hours of annotated first-person video data. - **Data Access**: Provides access to the dataset through command-line tools and a simple API for reading videos. - **Feature Extraction**: Offers lightweight wrappers for common models, including lightweight wrappers for models like Omnivore and SlowFast, which are essential for researchers and developers to analyze and utilize the dataset. - **Examples and Tutorials**: Includes a set of notebooks for Ego4D and Ego-Exo4D datasets, serving as examples and tutorials to help users understand and utilize the dataset effectively. ### Tech Stack: The repository is built using Python, with the `ego4d` module abstracting common video reading libraries such as `TorchAudio` and `PyAV`. ### Why It Matters: The Ego4d dataset and repository are crucial for researchers and developers in the field of video learning and perception

Overview: Macropodus is a comprehensive Natural Language Processing (NLP) toolkit built on top of the Albert+BiLSTM+CRF deep learning network architecture. It includes features such as Chinese word segmentation, part-of-speech tagging, named entity recognition, new word discovery, keyword extraction, text summarization, text similarity, scientific calculator, Chinese number to Arabic number conversion, Chinese to Roman number conversion, and pinyin conversion. Key features: - Chinese word segmentation (cut_bidirectional) - Part-of-speech tagging (cut_forward) - Named entity recognition (cut_reverse) - New word discovery - Keyword extraction - Text summarization - Text similarity - Scientific calculator - Chinese number to Arabic number conversion - Chinese to Roman number conversion - Pinyin conversion Tech stack: Python Why it matters: Macropodus is a powerful tool for natural language processing tasks, providing essential functionalities such as text summarization, keyword extraction, and more, making it highly useful for applications requiring NLP capabilities. Repository: yongzhuo/Macropodus Primary language: Python GitHub stars: 660 GitHub description: Macropodus is a comprehensive Natural Language Processing (NLP) toolkit built on top of the Albert+BiLSTM+CRF deep learning network architecture. It includes features such as Chinese word segmentation, part-of-speech tagging, named entity recognition, new word discovery, keyword extraction, text summarization, text similarity, scientific calculator, Chinese number to Arabic number conversion, Chinese to Roman number conversion, and pinyin

### Overview We formalize the randomness in diffusion models as a latent space, which can be inferred from real images. This allows us to perform zero-shot image-to-image translation with text-to-image diffusion models and traditional unpaired image-to-image translation using diffusion models trained on related domains. ### Key Features - **CycleDiffusion**: A simple method capable of zero-shot image-to-image translation with text-to-image diffusion models and traditional unpaired image-to-image translation using diffusion models trained on related domains. - **Latent Space Formalization**: We formalize the randomness in diffusion models as a latent space, which can be inferred from real images. - **Real Image Input**: All source images in our experiments are real images, including those generated by DALL•E 2. ### Tech Stack - Python ### Why It Matters This repository contains the code for CycleDiffusion, a method that leverages the latent space of diffusion models to perform zero-shot image-to-image translation and traditional unpaired image-to-image translation.

### Overview: IndoNLU is a comprehensive benchmark and resource for Indonesian natural language processing (NLP) that includes 12 downstream tasks, pre-trained models, and a starter code. This project was presented at AACL-IJCNLP 2020 and is designed to evaluate the performance of Indonesian language models. ### Key features: - **12 Downstream Tasks**: Covering various NLP tasks such as Named Entity Recognition, Part-of-Speech Tagging, Sentence Classification, and more. - **IndoBERT Models**: Large pre-trained models trained on a 4 billion word corpus, including two versions: IndoBERT-base and IndoBERT-lite. - **Indo4B Dataset**: A large pretraining dataset of 23 GB uncompressed and 5.6 GB compressed, used for model training and evaluation. - **GitHub Repository**: The project is hosted on GitHub, with 659 stars, making it a popular resource in the NLP community. - **Documentation**: Comprehensive documentation including contributing guidelines, examples, and submission instructions. ### Tech stack: Jupyter Notebook ### Why it matters: IndoNLU serves as a crucial benchmark for evaluating Indonesian language models, providing a standardized environment for researchers and practitioners to test and compare their models. Its inclusion of various downstream tasks and pre-trained models makes it an essential tool for advancing NLP research in Indonesia. ### Repository: IndoNLU - **Primary Language**: Jupyter Notebook - **GitHub Stars**: 659 - **GitHub Description**: The first-ever vast natural language processing

Overview: Polywise is an open-source agentic content system designed for self-hosted use, allowing users to interact with models, save knowledge, and retrieve context through a command-line interface or desktop application. Key features: - Command-line interface (CLI) for quick access - Desktop application for a seamless user experience - Model providers and integration setup - Embedding and rerank models for saved-content retrieval - Web UI for easy access to API - Authentication login for secure access Tech stack: TypeScript Why it matters: Polywise offers a flexible and accessible solution for creating and managing agentic content, making it ideal for developers looking to integrate AI into their projects without the need for a dedicated server. Repository: MatrixAges/polywise GitHub stars: 659

### Overview: EMA-Pytorch is a simple library designed to help developers keep track of an Exponential Moving Average (EMA) version of their PyTorch model. This allows for better performance monitoring and optimization of their models over time. ### Key Features: - **Exponential Moving Average (EMA) Tracking**: Automatically updates the EMA of your PyTorch model's weights. - **Post-Hoc Synthesis of EMA**: A method to create a new EMA model based on previously saved checkpoints, useful for post-hoc analysis. - **Target Representation Routing**: Automatically applies EMA outputs to the forward pass of nested module trees, such as those used in self-supervised learning. ### Tech Stack: Python ### Why It Matters: EMA-Pytorch simplifies the process of tracking and optimizing your PyTorch models by automatically updating their EMA versions. This can lead to better performance monitoring and optimization, making it easier to fine-tune your models over time.

### Overview: Awesome-LLM-Eval is a curated list of tools, datasets, benchmarks, demos, leaderboards, papers, and models primarily aimed at evaluating large language models (LLMs) and exploring the boundaries and limits of generative AI. ### Key features: - A comprehensive collection of tools, datasets, benchmarks, demos, leaderboards, papers, and models. - Focus on evaluation of LLMs, including various categories such as general, domain-specific, RAG-Evaluation, agent capabilities, coding capabilities, multimodal/cross-modal, long-context, inference speed, and quantization/compres - Updated regularly to reflect the latest developments in the field. ### Tech stack: Python, JavaScript, Markdown, and various open-source tools/libraries. ### Why it matters: This project serves as a valuable resource for researchers, developers, and practitioners interested in understanding and advancing the capabilities and limitations of large language models. By providing a curated list of resources, it facilitates the exploration of new evaluation methods and benchmarks, enabling deeper insights into the performance and potential of these models.

### Overview: InsightsLM is an open-source, self-hosted alternative to NotebookLM, designed to be a powerful AI research tool that can be customized and hosted privately. It uses Supabase and N8N as the backend and React as the frontend. ### Key Features: - **Chat with Your Documents:** Upload your documents and get instant, context-aware answers. - **Verifiable Citations:** Jump directly to the source of the information to ensure the AI isn't hallucinating. - **Podcast Generation:** Create audio summaries and discussions from your source materials, just like in NotebookLM. - **Private and Self-Hosted:** Maintain complete control over your data by hosting it yourself. Use local models if you wish. - **Customizable and Extensible:** Built with modern, accessible tools, making it easy to tailor to your specific needs. ### Tech Stack: - TypeScript ### Why It Matters: InsightsLM offers a powerful AI research tool that can be customized and hosted privately, making it a reliable window into your company's knowledge base. It bridges the gap between NotebookLM's closed-source nature and the potential for commercial opportunities in the commercialization of AI agents grounded in a company's specific knowledge.

Overview: AVA is a powerful AI assistant designed to enhance your Obsidian notes by integrating ChatGPT, making your note-taking experience more efficient and engaging. Key features: - Instantly format your notes with ChatGPT - Personalize your notes with AI-generated summaries and insights - Access AI-powered research and context for your notes - Collaborate with teammates by integrating with other tools Tech stack: TypeScript Why it matters: AVA leverages the cutting-edge capabilities of AI to streamline your Obsidian workflow, making it easier to organize, format, and collaborate on your notes. Its integration with ChatGPT ensures that your notes are always up-to-date and contextually relevant, enhancing your productivity and creativity.

### Overview: MEGABYTE is an implementation of the MEGABYTE model, a method for predicting sequences of large lengths using multiscale transformers in Pytorch. The model is designed to handle sequences up to a million bytes efficiently. ### Key Features: - **Multiscale Transformers**: The model uses transformers with different scales to handle sequences of various lengths. - **Generalization**: The implementation includes generalization to allow multiple local models. - **Pytorch**: The project is written in Pytorch, leveraging its capabilities for efficient computation and model training. ### Tech Stack: - Python ### Why It Matters: MEGABYTE is significant for its ability to predict sequences of large lengths, making it particularly useful in applications like language modeling and sequence prediction tasks where the input data can be extremely large. This project contributes to the advancement of transformer-based models by providing a practical implementation that can handle such data efficiently. ### Repository: lucidrains/MEGABYTE-pytorch - **Stars**: 658 - **Description**: Implementation of MEGABYTE, Predicting Million-byte Sequences with Multiscale Transformers, in Pytorch. - **Primary Language**: Python - **GitHub Description**: <img src="./MEGABYTE.png" width="450px"></img> ### Install: ```bash $ pip install MEGABYTE-pytorch ``` ### Usage: ```python import torch from MEGABYTE_pytorch import MEGABYTE model = MEGABYTE( num_tokens =

### Overview: Genv is an open-source GPU environment and cluster management system designed for ease of use and efficient GPU resource management. It simplifies the process of allocating and monitoring GPU resources, particularly in collaborative environments where data scientists and machine learning engineers need to share and manage GPUs effectively. ### Key Features: - **GPU Resource Control**: Easily manage and monitor GPU resources across machines or clusters. - **Team Collaboration**: Share GPUs among team members, enforce quotas, and switch between GPUs without code changes. - **Resource Pooling**: Pool GPUs from multiple machines for shared use. - **Open-Source LLM Support**: Serve and manage local open-source LLMs within the cluster for faster experimentation. ### Tech Stack: - Python - Open-source and community-driven ### Why It Matters: - **Ease of Use**: Simplifies GPU allocation and management for data scientists and ML engineers. - **Resource Control**: Ensures fair and efficient use of GPU resources. - **Collaboration**: Facilitates seamless sharing and management of GPUs among team members. - **Open-Source**: Free and easy to install, integrating with popular tools like Pyenv and Conda. ### GitHub Stars: 657 ### Repository: run-ai/genv ### Primary Language: Python ### GitHub Description: Genv is an open-source environment and cluster management system for GPUs, designed to ease the process of GPU allocation for data scientists and machine learning engineers without the need for code changes. ### Documentation: - [Documentation](https://docs.genv.dev/) - [Website](https://genv

### Overview: Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights is an official codebase for the paper "Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights" presented at ICML 2026. The project aims to explore the concept of "Neural Thickets," where multiple diverse task experts are densely distributed around a pretrained weight, providing a more comprehensive and diverse set of solutions. ### Key features: - **Codebase for Neural Thickets Paper**: The repository contains the official codebase for the paper "Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights" presented at ICML 2026. - **Iterative RandOpt Implementation**: Supports iterative RandOpt, an algorithm for optimizing neural networks. - **Support for Existing Frameworks**: Drops into existing Verl or Hugging Face TRL setups, making it easily integrable into existing projects. - **Diverse Task Expert Distributions**: The project aims to explore and demonstrate how diverse task experts can be densely distributed around a pretrained weight, providing a more comprehensive and diverse set of solutions. ### Tech stack: Python ### Why it matters: The project is significant as it provides a comprehensive and practical demonstration of the concept of Neural Thickets, showcasing how diverse task experts can be densely distributed around a pretrained weight. This can be particularly useful in scenarios where multiple expert opinions are needed for a more comprehensive solution, such as in recommendation systems, image classification, and other machine learning applications.

Overview: Mortiflix is a motion design studio that allows you to create videos step by step, with your approval at every stage. It uses Claude, a motion designer, to create the video and you to review each version. Key features: - Step-by-step video creation - Review and approval of each stage - Use of Claude Code or Anthropic API key - Pipeline management with `pipeline.json` and `PIPELINE.md` - Learning from your studio's style and taste Tech stack: JavaScript Why it matters: Mortiflix provides a more collaborative and detailed workflow compared to AI video tools that rely on a single prompt. It allows for real-time feedback and continuous improvement through user input, making the process more efficient and tailored to individual preferences.

Overview: Relax is an open-source reinforcement learning engine designed for post-training on large language models, leveraging asynchronous and multi-modal training capabilities. Key features: - Full Omni-Modal Training - Service-Oriented Architecture - Fully Async via TransferQueue - Hybrid Mode - Elastic Rollout Scaling - Rich Algorithm Suite Tech stack: Python Why it matters: Relax provides a robust framework for post-training reinforcement learning, enabling end-to-end multimodal training and efficient scaling of inference engines, making it suitable for large language model development and deployment.

Overview: Copulas is a Python library designed to model and generate multivariate data using copula functions. It allows users to choose from various univariate distributions and copulas, including Gaussian, Vine, and Archimedian, to create realistic synthetic datasets. Key features: - Model multivariate data - Compare real and synthetic data visually - Access and manipulate learned parameters Tech stack: Python Why it matters: Copulas offer a flexible and powerful way to model complex multivariate distributions, making it useful for various applications in data science, statistics, and machine learning. GitHub stars: 657 Repository: sdv-dev/Copulas Primary language: Python [](https://pypi.org/search/?c=Development+Status+%3A%3A+2+-+Pre-Alpha) [](https://pypi.python.org/pypi/copulas) [](https://pepy.tech/project/copulas) [](https://github.com/sdv-dev/Copulas/actions/workflows/unit.yml) [](

### Overview: Pgx is a collection of GPU-accelerated parallel game simulators designed for reinforcement learning (RL) in discrete state spaces. It leverages JAX for efficient parallel execution on accelerators, making it suitable for training models in environments like Chess, Shogi, and Go. ### Key features: - **Super fast performance**: Pgx excels in parallel execution on GPUs, providing high throughput. - **Wide variety of games**: Supports a range of games including Backgammon, Chess, Shogi, and Go. - **Beautiful visualization**: Outputs games in SVG format for easy visualization. ### Tech stack: - Python 3.9 - JAX - PyPI package ### Why it matters: Pgx is essential for researchers and developers looking to train reinforcement learning models in discrete state spaces, making it a valuable tool for improving AI in games and other applications. ### Repository: Repository: sotetsuk/pgx GitHub stars: 657 Primary language: Python License: Apache 2.0 CI: GitHub Actions Codecov: GitHub Actions Arxiv: 2303.17503 Docs: Read the Docs GitHub Actions: ci.yml Colab: pgx_hello_world.ipynb, baselines.ipynb, pgx2pettingzoo.ipynb Training examples: AlphaZero, PPO ### Quick start: - **Getting started**: Explore the Colab notebook for a quick introduction. - **Baselines**: Learn from the baselines notebooks for

Overview: Dexto is an agent harness designed to turn LLMs into reliable, stateful AI agents. It includes a powerful coding agent and tools for building various applications like autonomous agents, digital companions, and multi-agent systems. Key features: - **Coding Agent**: A production-ready coding agent for building and managing AI applications. - **Autonomous Agents**: Plan, execute, and adapt to user goals. - **Digital Companions**: Assistants that remember context and anticipate needs. - **MCP Clients & Servers**: Connect tools, files, and APIs via Model Context Protocol. - **Multi-Agent Systems**: Agents that collaborate and solve complex tasks together. Tech stack: TypeScript Why it matters: Dexto provides a streamlined solution for building AI applications, making it easier to configure and use different LLMs, and offering tools for managing and collaborating with multiple agents. Repository: truffle-ai/dexto Primary language: TypeScript GitHub stars: 657 GitHub description: Agent harness and tookit for building AI agents and agentic applications. CLI and SDKs included README excerpt: <a href="https://dexto.ai">Dexto</a> is an agent harness designed to turn LLMs into reliable, stateful AI agents. It includes a powerful coding agent and tools for building various applications like autonomous agents, digital companions, and multi-agent systems. The coding agent is just one example of what you can build. Create your own agents by defining a YAML config—same architecture, your domain. Quick Start: Install via

### Overview: Exclusively Dark (ExDARK) is a comprehensive low-light image dataset designed for object detection and image enhancement research. Comprising 7,363 images captured in various low-light conditions, including 10 different environments, ExDARK provides detailed annotations for both image classes and local object bounding boxes, making it a valuable resource for researchers. ### Key Features: - **Image Class and Object Level Annotations:** Detailed annotations for both image classes and local object bounding boxes. - **10 Different Environments:** Images captured in 10 distinct low-light conditions. - **7,363 Images:** A total of 7,363 images for extensive research and benchmarking. - **MATLAB Code:** The source code for low-light image enhancement is available, facilitating further research. ### Tech Stack: MATLAB ### Why It Matters: Exclusively Dark (ExDARK) is a crucial dataset for researchers focusing on low-light image processing and object detection. Its comprehensive nature, including detailed annotations and diverse image conditions, makes it an indispensable resource for advancing the field of low-light imaging and image enhancement. ### Repository: Repository: cs-chan/Exclusively-Dark-Image-Dataset GitHub Stars: 654 GitHub Description: The Exclusively Dark (ExDARK) dataset is a collection of 7,363 low-light images from very low-light environments to twilight (i.e., 10 different conditions) with 12 object classes (similar to PASCAL VOC) annotated on both image

### Overview: Vestige is a local, deterministic safety kernel designed for AI agents. It ensures that every memory write and agent action is recorded in Strata, an append-only log, allowing for precise tracking and review of actions. ### Key Features: - **Guard against force pushes, dropped tables, and production deploys** using Operator Lite. - **Backfill mechanism** that remembers what matters, allowing agents to see and correct their past actions. - **Strata log** for precise memory tracking and retrieval. - **Operator Lite** guard that stops unapproved actions and provides a weekly report. ### Tech Stack: Rust ### Why It Matters: Vestige ensures that AI agents are responsible and accountable for their actions, preventing accidental or malicious changes to their environment. It provides a clear and verifiable history of agent behavior, making it easier to understand and debug complex systems. ### Repository: GitHub Repository: samvallad33/vestige ### GitHub Stars: 656 ### Primary Language: Rust ### Binary Platforms: - macOS - Linux ### Release Targets: - 6 ### License: AGPL-3.0 ### Quick Start: 1. Install the binaries or download them from Releases. 2. Run `npx @vestige/init` to choose between Memory or Memory + Guard. 3. Restart your AI client and say: *"Tell me something about a friend."* Ask about them later. ### Backfill Mechanism: - **Retroactive Salience Backfill**: When a new memory makes an old one