Research & Education

Lm-evaluation-harness is a framework specifically designed for the few-shot evaluation of language models. It provides a robust environment for researchers and engineers to assess the performance of different models across a variety of tasks. The tool is built with a focus on usability, offering CLI refactoring with subcommands and support for YAML configuration files. Additionally, it provides lighter installation options through separate model backends, making it more flexible for different setups.

OpenADMET ExpansionRx Blind Challenge provides a collaborative platform focused on advancing ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) property prediction for drug discovery. Participants download provided training data, develop and train their own predictive models, and then submit their predictions in CSV file format. This initiative aims to foster open-source research and development in the critical area of ADMET, which is crucial for identifying viable drug candidates.

notebooks provides a comprehensive collection of computer vision tutorials designed to educate users on cutting-edge models and techniques. It delves into advanced architectures such as ResNet, YOLOv11, and SAM, offering practical insights into their implementation and application. The resource is particularly useful for individuals and teams working on computer vision challenges, including object detection, image segmentation, and pose estimation tasks. It aims to equip users with the knowledge to understand and apply complex computer vision concepts.

open_spiel is a comprehensive framework designed for research in reinforcement learning within the context of games. It offers a robust collection of environments and algorithms, facilitating the exploration of general reinforcement learning and advanced search/planning techniques. The framework is versatile, supporting a wide array of game structures, including n-player zero-sum, cooperative, and general-sum games. It is also adaptable for both one-shot and sequential game scenarios, making it a valuable tool for researchers and developers in the field.

BIG-bench is an AI benchmarking platform specifically designed to evaluate and enhance the performance of various AI models. It provides a comprehensive testing suite, making it a valuable resource for both AI researchers and developers. As an open-source platform, BIG-bench actively promotes collaboration and innovation within the AI community, continuously evolving its repository of AI benchmarks. The platform is notable for containing over 200 distinct tasks, offering a wide range of evaluation scenarios.

bi-att-flow is an implementation of the Bi-Directional Attention Flow (BiDAF) network, specifically designed for machine comprehension tasks. This network excels at understanding and processing text by representing context at various levels of granularity. A core feature is its bi-directional attention flow mechanism, which enables a query-aware context representation, allowing the model to effectively focus on relevant parts of the text based on a given query. This makes it suitable for applications requiring deep textual understanding.

Awesome-World-Model is a comprehensive, curated list specifically focused on World Models relevant to Autonomous Driving and Robotics. This resource is designed for researchers and practitioners in the AI field, providing a centralized location to discover, track, and benchmark the latest World Model methodologies. It also includes a survey of the field, offering valuable context and insights into the current state of World Model research and applications.

bd3lms is a project focused on Block Diffusion, an innovative method that bridges the gap between autoregressive and diffusion language models. This research was recognized with an oral presentation at ICLR 2025, highlighting its significance in the field of AI. The project serves as a central hub for resources and detailed information pertaining to this advanced language model interpolation technique, catering to researchers and academics interested in the latest developments in AI.

contrastive-predictive-coding is a Keras-based tool that implements the Representation Learning with Contrastive Predictive Coding algorithm. Its primary function is to learn meaningful data representations by capturing semantic information without the need for explicit annotations. The tool leverages unsupervised learning methods to identify and recognize patterns within data, making it a valuable resource for advancing AI research and development. It is designed for those looking to explore and apply advanced representation learning techniques.

CV-pretrained-model offers a collection of pre-trained computer vision models, designed to provide a significant head start for various computer vision tasks. Instead of building models from scratch, users can leverage these existing models as a foundation for similar problems. While not guaranteed to be 100% accurate for every specific use case, these pre-trained models offer a robust starting point, saving considerable time and resources in the development process. This repository is ideal for those looking to quickly implement or experiment with computer vision solutions.

D-NeRF is a technique designed for generating new perspectives of scenes that are in motion. It leverages neural radiance fields (NeRF) to create a comprehensive representation of dynamic environments. This allows users to render these scenes from any viewpoint and at any specific moment in time. A key capability of D-NeRF is its ability to effectively manage and represent complex geometries that are non-rigid, making it suitable for a wide range of dynamic visual applications.

train-deepseek-r1 is a project dedicated to the ground-up construction of DeepSeek R1 models. It leverages reinforcement learning, building upon the DeepSeek V3 base model. The project emphasizes ease of use, providing flowcharts and detailed step-by-step implementation guides to streamline the training process. Its core functionality allows users to develop their own custom models utilizing the tinygrad framework, making advanced AI model creation more accessible.

Awesome-Computer-Vision-Paper-List is a curated repository specifically designed for computer vision researchers. It compiles papers that have been accepted at leading AI conferences, providing a centralized resource for academic exploration. Users can efficiently search for papers based on specific research areas, streamlining the process of literature review. The primary goal of this tool is to assist researchers in conveniently locating relevant academic work and keeping abreast of the most recent developments and breakthroughs within the dynamic field of computer vision.

Awesome-hand-pose-estimation is a comprehensive, curated list of resources specifically focused on hand pose estimation and tracking. This valuable collection includes direct links to a variety of essential materials, such as evaluation datasets, arXiv papers, journal papers, and conference papers. It serves as a central hub for researchers and developers who are actively engaged in the field of hand pose estimation, offering easy access to foundational and cutting-edge research.

Awesome-Scientific-Language-Models provides a comprehensive, curated list of pre-trained language models tailored for various scientific domains. This resource is designed to assist researchers and developers who are actively working with language models in scientific applications, offering a centralized collection of relevant tools and models. The repository is open-source, encouraging community contributions to keep the list updated and expansive, thereby fostering collaboration within the scientific AI community.

PhoGPT is a generative pre-trained model tailored for the Vietnamese language, featuring both a base model (PhoGPT-4B) and a chat variant (PhoGPT-4B-Chat). Both models are equipped with 3.7 billion parameters, indicating a substantial capacity for language processing. The base model has undergone pre-training on an extensive Vietnamese corpus, enabling it to understand and generate Vietnamese text effectively. PhoGPT's primary objective is to foster advancements in Vietnamese language AI research and its practical applications.

Anatomy of BoltzGen offers a detailed exploration of the architecture and design principles behind BoltzGen. This resource provides a deep dive into the system's various components and their structural relationships. It is specifically designed for educational purposes, helping users understand the intricate inner workings of BoltzGen. AI researchers can also leverage this tool to gain comprehensive insights into the system's design.

Awesome-vlm-architectures is a comprehensive, curated list focusing on Vision-Language Models (VLMs) and their underlying architectures. VLMs are designed to process both image and text data concurrently, facilitating advanced AI tasks such as Visual Question Answering (VQA) and automated image captioning. The repository serves as a valuable resource for researchers and developers interested in exploring and understanding the intricacies of multimodal fusing and masked-language modeling techniques within the VLM domain.

Awesome-tiny-object-detection is a comprehensive, curated list specifically designed for researchers and developers interested in the field of tiny object detection. This resource compiles a wide array of academic papers and related materials, covering various sub-topics such as general tiny object detection, tiny face detection, and tiny pedestrian detection. Beyond just papers, the list also includes links to relevant datasets, in-depth surveys, and informative articles, making it a central hub for discovering and accessing key resources in this niche area of computer vision.

CVA6 is a sophisticated 6-stage RISC-V core, engineered for both application and embedded system development. It offers high configurability, allowing it to be adapted to various project requirements. A key feature is its ability to boot Linux in application configurations, highlighting its robustness for complex operating environments. The core strictly adheres to the 64-bit RISC-V instruction set architecture and is structured as a single-issue, in-order CPU, providing a clear and efficient processing pipeline for developers.

cv-arxiv-daily is a tool designed to streamline the process of tracking new research in computer vision. It automatically updates a curated list of papers daily, leveraging GitHub Actions for this process. The tool provides users with direct links to PDFs and associated code, making it easier for researchers and AI enthusiasts to access and review the latest publications in their field. Its primary goal is to keep its audience informed about new advancements without manual tracking.

DINOv2 is a self-supervised learning framework implemented in PyTorch, designed to facilitate various computer vision applications. It provides researchers and developers with pre-trained models and codebases, enabling them to leverage self-supervised learning techniques without extensive manual labeling. The tool specifically mentions support for loading XRay-DINO backbones, suggesting potential applications in medical imaging, and Channel-Adaptive DINO code, indicating flexibility in handling different data modalities or architectures. Its focus on providing readily available components aims to accelerate development in computer vision.

KOFFVQA Leaderboard is an AI tool specifically designed for benchmarking and evaluating Visual Question Answering (VQA) models. It provides a platform for researchers and engineers to compare the performance of various AI models against each other using the KOFFVQA dataset. The tool's primary purpose is to facilitate the tracking of progress within the VQA field and to identify top-performing models, thereby aiding in the advancement of VQA technology.

DDAD is a specialized dataset developed for advancing autonomous driving research. Its primary focus is to provide dense depth information, which is crucial for accurate long-range depth estimation, particularly in complex urban environments. The dataset is comprehensive, offering detailed sensor placement information and predefined evaluation metrics to facilitate standardized research and development. It is a valuable resource for researchers and engineers working on perception systems for autonomous vehicles.