Research & Education

WASP (Wallenberg AI, Autonomous Systems and Software Program) is Sweden's largest individual research program, dedicated to fostering excellence in artificial intelligence, autonomous systems, and software. It supports strategically motivated basic research, provides extensive education through its Graduate School, and focuses on faculty recruitment to strengthen national competence. The program offers unique opportunities for achieving international research excellence with industrial relevance, funded by the Knut and Alice Wallenberg Foundation. WASP emphasizes interdisciplinary collaboration, offering PhD programs, industrial cooperation, and various events to connect researchers and industry professionals. Its vision is to benefit Swedish industry through cutting-edge research and competence development.

MADRL is a repository offering code for multi-agent deep reinforcement learning (MADRL), providing implementations of several multi-agent reinforcement learning environments. These include Pursuit Evasion, Waterworld, Multi-Agent Walker, and Multi-Ant. The package requires OpenAI Gym and a forked version of rllab (the multiagent branch) for its functionality. It is designed for researchers and developers in the field of multi-agent reinforcement learning, allowing them to set up and run simulations with curriculum learning. The project also provides details on policy definitions and includes a citation for its accompanied paper, making it a valuable resource for academic and practical applications in MADRL.

MaxText is a high-performance, highly scalable, open-source library for Large Language Models (LLMs), implemented in pure Python/JAX. It is designed to run efficiently on Google Cloud TPUs and GPUs, supporting both pre-training and scalable post-training with techniques like Supervised Fine-Tuning (SFT) and Reinforcement Learning (GRPO, GSPO). MaxText achieves high Model FLOPs Utilization (MFU) and tokens/second across various cluster sizes, leveraging the power of JAX and the XLA compiler. It offers a library of high-performance models including Gemma, Llama, DeepSeek, Qwen, and Mistral, and serves as a launching point for ambitious LLM projects in research and production. Users can experiment with MaxText out of the box or fork and modify it to meet specific needs, with support for multi-modal training.

ELLIS (European Laboratory for Learning and Intelligent Systems) is a pan-European AI network dedicated to strengthening Europe's leadership in AI through research excellence. Founded in 2018, ELLIS unites top AI researchers across Europe and Israel, focusing on machine learning as the driver for modern AI. The network establishes multi-centric AI research laboratories, including ELLIS Units and Institutes, distributed across 17 countries. It also runs 16 pan-European Research Programs that push scientific boundaries in areas from theoretical foundations to applications in health and climate sciences. Furthermore, ELLIS offers a PhD & Postdoc Program, pairing outstanding students with leading researchers for international exchanges and joint supervision, cultivating the next generation of AI talent.

Florence2 + SAM2 is an advanced AI tool designed for precise image and video segmentation. It leverages the capabilities of Florence2 and SAM2 models to enable users to upload visual content and define objects of interest through text prompts. The application then highlights these specified objects, providing detailed segmentations, captions, or descriptions. This makes it particularly useful for tasks requiring accurate object identification and isolation within complex visual data. While the tool is currently experiencing a runtime error, its intended functionality targets AI researchers, developers, and professionals in image processing who need robust segmentation capabilities.

MAmmoTH2 is presented as the strongest open-source large language model (LLM) specifically designed for reasoning tasks. Hosted on Hugging Face Spaces by TIGER-Lab, this tool allows users to interact with a chatbot that generates text responses based on conversation history and a provided system prompt. It is primarily intended for research and development, enabling users to explore advanced language understanding and complex problem-solving. As an open-source model, MAmmoTH2 offers flexibility for developers and researchers to integrate and adapt it for various applications.

The Machine Learning Problem Bible (MLPB) is an open-source repository designed to be an organized collection of machine learning problems and their corresponding solutions. It addresses the common challenge faced by machine learning practitioners: finding relevant examples to adapt for their own projects. MLPB allows users to browse problems based on specific tags like "multi-class classification," "sparse-data," or "NLP," making it easy to locate solutions for various scenarios. Each problem typically includes a dedicated data directory and one or more scripts demonstrating a solution, often in both R and Python. The repository aims to provide practical examples, such as comparing different machine learning models or predicting ranked target variables, thereby serving as a valuable resource for learning and problem-solving in machine learning.

This GitHub repository, 'state-of-the-art-result-for-machine-learning-problems,' serves as a comprehensive collection of state-of-the-art (SoTA) results across numerous machine learning domains. It covers supervised learning (Speech, Computer Vision, NLP), semi-supervised learning (Computer Vision), unsupervised learning (Speech, Computer Vision, NLP), transfer learning, and reinforcement learning. The repository aims to stay current by actively soliciting contributions and issue reports from the community, allowing users to submit new SoTA results via a Google Form or GitHub issues. Each entry typically includes the research paper name, dataset, metric, source code, and year, making it a valuable resource for researchers and practitioners to benchmark models and stay informed on the latest advancements in the field.

Open CoT Leaderboard is a platform designed to track, rank, and evaluate the Chain-of-Thought (CoT) quality of open large language models (LLMs). Hosted as a Hugging Face Space, it provides a centralized location for researchers and developers to browse and filter a leaderboard of LLM benchmarks. Users can submit their own models for evaluation, allowing for comparison against existing models and contributing to the collective understanding of LLM performance. The platform offers transparency into the evaluation process and the status of submitted models, making it a valuable resource for identifying top-performing open-source LLMs and advancing AI research.

The Open Japanese LLM Leaderboard is a platform designed for exploring and comparing large language models (LLMs) tailored for the Japanese language. Hosted on Hugging Face Spaces, this tool allows users to search for models by name, and apply filters based on type, size, and precision. It provides performance metrics and visualizations to help researchers, developers, and enthusiasts assess the capabilities of various Japanese LLMs. The leaderboard aims to facilitate the identification of top-performing models, supporting advancements in Japanese natural language processing and AI development. While the current live website indicates a runtime error, the intended functionality is to offer a comprehensive resource for evaluating and understanding the landscape of open Japanese LLMs.

The Artificial Superintelligence Alliance (ASI) unites Fetch.ai, SingularityNET, and CUDOS to advance decentralized Artificial General Intelligence (AGI) and Artificial Superintelligence (ASI). It offers a robust, open-source innovation stack, including ASI:Chain, an AI-native Layer-1 blockchain, and ASI:Cloud, a permissionless GPU cloud for AI inference. The alliance also develops ASI-1 Mini, a Web3-native LLM, and conducts research into advanced AGI architectures like OpenCog Hyperon and the MeTTa programming language. Its mission is to empower developers, enterprises, and researchers to build ethical, scalable, and groundbreaking AI solutions, ensuring advanced intelligence remains a shared, accessible resource.

AIARD (Artificial Intelligence Assisted R&D) develops AI tools specifically designed to assist users in the research and development process. The platform focuses on helping users articulate complex problems and efficiently locate pertinent scientific information. AIARD aims to revolutionize how companies approach problem formulation and resolution by integrating AI and TRIZ methodologies. A core capability of the tool involves extracting and digitizing knowledge from various documents, streamlining the information gathering and analysis phases of R&D. This approach helps users to quickly access and leverage existing knowledge, fostering innovation and efficiency in their research endeavors.

gantts offers a PyTorch implementation for Generative Adversarial Networks (GAN) based text-to-speech (TTS) and voice conversion (VC). This open-source project allows developers and researchers to experiment with advanced speech synthesis techniques. Key features include the ability to generate audio samples, configure hyper-parameters for fine-tuning speech quality, and integrate with various datasets like CMU ARCTIC. The tool provides scripts for acoustic feature extraction, linguistic/duration feature extraction, and GAN-based training, making it suitable for both TTS and VC model development. It also includes evaluation scripts for both applications and supports monitoring training progress via TensorBoard.

The Center for Human-Compatible AI (CHAI) is a research institute based at UC Berkeley dedicated to ensuring artificial intelligence systems are provably beneficial for humanity. CHAI's core mission involves developing the conceptual and technical frameworks necessary to guide AI research towards human-compatible outcomes. Their work includes exploring topics like offline reinforcement learning, defining political neutrality for AI, and investigating computational frameworks for human care. They also address fundamental coordination problems, such as learning to yield and request control in AI systems. CHAI publishes research, hosts a blog, and offers opportunities for faculty, staff, researchers, and students to contribute to their mission.

CeADAR Ireland is the national center for Applied AI in Ireland, funded by Enterprise Ireland and IDA. Its core mission is to assist businesses and organizations in exploring, experimenting with, and integrating innovative AI solutions into their processes and products. This aims to enhance productivity, competitiveness, digitalization, and sustainability. CeADAR offers independent, unbiased advice and develops advanced AI solutions without being tied to any specific technology stack. It also helps companies secure funding, boasts deep technical expertise across all AI and ML domains, and operates as a designated European Digital Innovation Hub (EDIH), providing 100% discounted services to eligible enterprises and public service organizations in Ireland. Additionally, CeADAR builds partnerships and consortiums leveraging its extensive EU-wide network.

The Open Portuguese LLM Leaderboard provides a comprehensive platform for tracking, ranking, and evaluating open Large Language Models (LLMs) specifically designed for the Portuguese language. Users can easily explore and filter models based on various criteria such as type, size, precision, and language. This tool is invaluable for researchers, developers, and AI enthusiasts who need to compare the performance of different LLMs in Portuguese. By offering detailed benchmarks, it helps identify top-performing models for specific Portuguese language tasks, facilitating informed decision-making in model selection and development. The platform aims to foster innovation and collaboration within the Portuguese AI community by providing transparent and accessible performance metrics.

Qwen2.5-Math represents a specialized series of large language models from the Qwen2 family, specifically engineered to excel in mathematical problem-solving and research. These models are tailored to handle complex mathematical queries, equations, and theoretical concepts, providing advanced capabilities for users in academic and scientific fields. By focusing on mathematics, Qwen2.5-Math aims to offer more accurate and relevant solutions compared to general-purpose LLMs. The models are accessible through popular platforms like Hugging Face and ModelScope, facilitating integration and experimentation for researchers and developers working on AI-driven mathematical applications.

regl-cnn is an open-source project designed for GPU-accelerated handwritten digit recognition, leveraging Convolutional Neural Networks (CNNs) within WebGL. This tool serves as a practical demonstration of how to implement a CNN directly on the GPU using WebGL, offering insights into high-performance computing for machine learning in web environments. The underlying network was initially trained using TensorFlow, and subsequently, its architecture and functionality were meticulously reimplemented in WebGL to showcase client-side inference capabilities. It is particularly useful for web developers interested in integrating machine learning models into web applications and machine learning enthusiasts looking to understand GPU-accelerated CNNs.

pytorch-openai-transformer-lm offers a PyTorch implementation of OpenAI's finetuned transformer language model, based on the paper "Improving Language Understanding by Generative Pre-Training." This tool includes a script to import the weights pre-trained by OpenAI, allowing users to leverage the model within a PyTorch environment. It supports fine-tuning the pre-trained model for classification tasks, with an example provided for the ROCStories Cloze task. The implementation closely follows the original TensorFlow code, including a modified Adam optimization algorithm with fixed weight decay and scheduled learning rate. It provides classes for a full language model with a tied decoder and a classifier head on top of the transformer.

The Big Data for Smart Society Institute (GATE) is a Centre of Excellence in Bulgaria, established in 2019 as an autonomous structure of Sofia University “St. Kliment Ohridski”. GATE focuses on integrating and extending scientific excellence and innovation in priority areas such as Big Data and Artificial Intelligence. The institute is dedicated to attracting, inspiring, and cultivating the next generation of Early-Stage Researchers, guiding them in the fields of Big Data and AI. Research at GATE is concentrated on four main areas: Data management, Data analytics, Data insight, and Data engineering. GATE also engages in collaborative R&D projects, contract research, education, and training, and works on creating spin-outs and start-ups.

slimevolleygym is an OpenAI Gym environment designed for testing single and multi-agent reinforcement learning algorithms through a simple Slime Volleyball game. This environment is lightweight, requiring only gym and numpy as dependencies, making it less prone to breaking and easy to integrate. It features a baseline 120-parameter neural network opponent, which can be replaced for multi-agent or self-play scenarios. The environment runs efficiently, achieving around 12.5K timesteps per second on state-space observations, facilitating faster iteration in experiments. It supports both state-space and pixel observations, with the latter mimicking Atari Learning Environment setups, and includes a tutorial for various training methods. The environment is particularly useful for educational purposes and for exploring advanced RL methods like self-play and continual learning.

SimpleTuner is a comprehensive, open-source fine-tuning kit designed for image, video, and audio diffusion models. It prioritizes simplicity and code understandability, making it an ideal academic exercise and collaborative development platform. The tool features a user-friendly web UI, multi-modal and multi-GPU training capabilities, and advanced caching for faster training. It supports various model architectures, including Stable Diffusion XL, Stable Diffusion 3, and Flux, with integrations for DeepSpeed and FSDP2 for memory optimization. SimpleTuner also includes enterprise-grade features like worker orchestration, SSO integration, role-based access control, and a job queue with priorities, all available for free.

simple-neural-network is a Python script designed to illustrate the backpropagation algorithm, a fundamental concept in neural network training. This open-source tool serves as an educational resource for individuals interested in the inner workings of artificial neural networks. It provides a clear, step-by-step example of how neural networks learn by adjusting weights based on error signals. The script is particularly useful for students, AI enthusiasts, and developers who want to gain practical insight into the backpropagation process without needing to build a complex neural network from scratch. Its simplicity makes it an accessible entry point for understanding more advanced machine learning concepts.

SkyRL is a modular, open-source, full-stack reinforcement learning (RL) library specifically designed for large language models (LLMs). It aims to streamline research and development in the field of AI agents by offering a flexible framework for building and training intelligent agents. While the provided website content is a GitHub pricing page for GitHub itself, the tool's description indicates its core purpose is to support advanced AI development. Researchers and developers can leverage SkyRL to experiment with and implement various RL algorithms tailored for LLM applications, fostering innovation in AI agent capabilities and performance.