Research & Education

PromptBench is a PyTorch-based Python package designed as a unified evaluation framework for large language models (LLMs). It offers user-friendly APIs for researchers and developers to conduct comprehensive evaluations of LLMs, including quick performance assessments, prompt engineering method testing (like Chain-of-Thought, Emotion Prompt, and Expert Prompting), and adversarial prompt robustness analysis. The framework integrates dynamic evaluation techniques such as DyVal to mitigate test data contamination and efficient multi-prompt evaluation with PromptEval. It supports a wide range of language and multi-modal datasets and models, both open-source and proprietary, making it a versatile tool for understanding and benchmarking LLM capabilities.

StableDiffusionReconstruction is a research-oriented tool designed for reconstructing visual experiences directly from human brain activity. Utilizing Stable Diffusion models, it allows for the generation of high-resolution images based on neural data. The project, stemming from research by Takagi and Nishimoto presented at CVPR 2023, also incorporates advanced decoding techniques. These include methods for decoding text prompts from brain activity, integrating GANs for improved image quality, and incorporating decoded depth information, significantly enhancing reconstruction accuracy. This repository provides the necessary code and instructions for reproducing these methods, making it a valuable resource for researchers in neuroscience and AI.

Perturbed-Attention Guidance SDXL is an AI tool designed for image generation, leveraging the power of Stable Diffusion XL models with a unique perturbed attention guidance mechanism. This innovative approach enables users to produce distinctive and artistic images. The application presents two side-by-side results, with the left image showcasing the perturbed attention guidance technique. While the tool was previously available as a Hugging Face Space, it is currently paused. Users interested in utilizing this Space are encouraged to reach out to the author(s) via the community tab to request its restart.

Speech-Emotion-Recognition is an open-source project designed for identifying emotions in spoken language. It leverages various machine learning models, including Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), Support Vector Machines (SVM), and Multilayer Perceptrons (MLP), all implemented within the Keras framework. The tool focuses on advanced feature extraction techniques, which contribute to its reported accuracy of around 80%. It supports Python and integrates with essential libraries such as scikit-learn for model training and evaluation, and librosa for audio feature processing. This makes it a valuable resource for researchers and developers working on speech analysis and emotion detection applications.

Show-1 is an advanced open-source text-to-video generation model developed by Show Lab at the National University of Singapore. It uniquely combines pixel and latent diffusion models to create videos from textual descriptions. The tool provides access to various model weights, including a base model, an interpolation model, and super-resolution models, which can be downloaded from HuggingFace. Users can generate videos by running a Python script, with outputs saved in GIF format. Show-1 also offers a Gradio demo for local use and has been accepted to IJCV, highlighting its academic recognition. It is designed for researchers and developers interested in cutting-edge video synthesis.

Static-to-Dynamic-LLMEval is the official GitHub repository for a paper detailing recent advances in large language model benchmarks, specifically focusing on data contamination. The project conducts an in-depth analysis of existing static-to-dynamic benchmarking methods designed to reduce data contamination risks. It examines methods that enhance static benchmarks, identifies their limitations, and highlights the critical gap in standardized criteria for evaluating dynamic benchmarks. The repository proposes optimal design principles for dynamic benchmarking and analyzes the limitations of current dynamic benchmarks, offering a comprehensive overview of advancements in data contamination research and guiding future efforts.

Technical_Book_DL is a comprehensive technical book on deep learning, offering a pedagogical approach to understanding the three most common neural network architectures: Feedforward, Convolutional, and Recurrent. For each architecture, the book meticulously details its fundamental building blocks. It then proceeds to derive the forward pass and the complete update rules for the backpropagation algorithm, providing a thorough understanding for students and AI enthusiasts. The entire document is available as a downloadable PDF, with all figures and LaTeX source files also provided in the repository for compilation. This resource is particularly valuable for those who prefer detailed, indexed formulas over abstract matrix formulations, ensuring a precise grasp of the underlying mechanics.

system-prompts-and-models-of-ai-tools is a comprehensive open-source GitHub repository that curates system prompts, internal tools, and AI models from a wide array of AI applications. This resource is invaluable for developers, researchers, and AI enthusiasts looking to understand the underlying mechanics and prompt engineering strategies of popular tools like Augment Code, Claude Code, Cursor, Devin AI, NotionAI, Perplexity, and many others. It provides a centralized location to explore how different AI systems are structured and prompted, fostering learning and innovation in the AI development community. The repository also highlights the importance of securing AI systems against prompt injection and extraction risks.

TabM is an official open-source repository for the paper "TabM: Advancing Tabular Deep Learning With Parameter-Efficient Ensembling" (ICLR 2025). It offers a PyTorch-based Python package for implementing the TabM model, along with layers and tools for constructing custom architectures that efficiently ensemble MLP-like models. The tool is designed to improve performance on challenging tabular benchmarks like TabReD and has been successfully applied in Kaggle competitions. TabM is noted for its efficiency, being faster than prior tabular deep learning methods and capable of handling large datasets up to 100M+ objects. It allows for parallel training and weight sharing among MLPs, leading to better runtime, memory efficiency, and task performance.

TrajectoryCrafter is an advanced Content & Design tool designed to redirect camera trajectories in monocular videos using sophisticated diffusion models. This tool, presented at ICCV 2025, allows users to generate high-fidelity novel views from standard monocular video footage, offering precise control over camera pose. It is particularly useful for researchers and developers working with video manipulation and synthesis. The system requires a GPU with at least 28GB VRAM for optimal performance and can be set up using standard Python environments. While powerful, its capabilities are rooted in a pretrained video diffusion model, meaning it performs best with well-defined objects and clear motion, and may face limitations with highly complex scenarios beyond its base model's generation capacity. It provides both command-line inference and a local Gradio demo for ease of use.

Trending-Deep-Learning is a GitHub repository that provides a curated list of the top 100 trending deep learning projects. This resource is updated regularly and sorts repositories based on the number of stars they gained on a specific day. It leverages the GitHub search API with a comprehensive query including terms like 'deep-learning', 'CNN', 'RNN', 'convolutional neural network', and 'recurrent neural network'. Repositories with 40,000 stars or more are excluded to focus on emerging trends. This tool is ideal for researchers, developers, and students looking to stay updated on the latest advancements and popular projects within the deep learning community, offering a quick overview of what's gaining traction.

Urban-Sound-Classification is an open-source deep learning project designed for the classification of urban sounds. It offers a comprehensive set of Jupyter notebooks demonstrating various neural network architectures, including feedforward, convolutional, and recurrent neural networks. The project is built using Python 3.5 (or above) and leverages popular libraries such as Tensorflow 2.x, Numpy, Matplotlib, and Librosa. It primarily uses the UrbanSound8k dataset for model training, with Google's AudioSet suggested as an alternative. This tool is ideal for researchers, students, and developers interested in deep learning applications for audio analysis and sound classification, providing a practical foundation for understanding and implementing these techniques.

Türkçe Yapay Zeka Kaynakları is a comprehensive, open-source repository dedicated to deep learning and machine learning resources available in Turkish. Supported by the Deep Learning Türkiye community, this platform centralizes a wide array of materials including blog posts, video lectures, scientific articles, code examples, and datasets. It serves as an invaluable hub for individuals seeking to learn or conduct research in AI within the Turkish language. The resource is continuously updated and encourages contributions from the community, ensuring a rich and current collection of information across various AI topics, algorithms, frameworks, and applications.

XuanCe (玄策) is an open-source, comprehensive, and unified deep reinforcement learning (DRL) library designed to provide high-quality and easy-to-understand implementations of DRL algorithms. It aims to address the sensitivity of DRL algorithms to hyper-parameter tuning and unstable training processes by offering a robust and flexible framework. XuanCe is highly modularized, easy to install and use, and supports flexible model combinations. It includes abundant algorithms for various tasks, supporting both DRL and Multi-Agent Reinforcement Learning (MARL) tasks. The library boasts high compatibility across different deep learning backends (PyTorch, TensorFlow2, MindSpore), operating systems (Linux, Windows, MacOS), and hardware (CPU, GPU). Key features include fast running speed with parallel environments, distributed training with multi-GPUs, automatic hyperparameter tuning, and good visualization effects with TensorBoard or Weights & Biases.

ViewCrafter is an open-source research project designed for high-fidelity novel view synthesis, leveraging advanced video diffusion models. It enables users to generate novel views from either a single reference image or sparse reference images, offering highly precise control over camera pose. The tool is associated with a paper published in TPAMI 2025, highlighting its academic rigor and cutting-edge capabilities in computer vision. ViewCrafter provides pretrained models for various resolutions and frame rates, catering to both single-view and sparse-view novel view synthesis tasks. It includes scripts for inference, evaluation, and even a local Gradio demo, making it accessible for researchers and developers to experiment with and integrate into their workflows. The project emphasizes its research exploration nature, acknowledging potential variability in results due to the underlying video diffusion models.

Visualizer is a specialized tool designed to simplify the process of visualizing attention maps within deep learning models, particularly those based on Transformer architectures. It addresses common challenges faced by developers, such as the difficulty of extracting deeply nested attention maps without modifying model code or encountering out-of-memory errors. The tool provides a non-intrusive method using Python decorators and PyTorch hooks, allowing users to precisely retrieve intermediate variables like attention maps. This ensures consistency between training and testing phases, as no code changes are required for visualization. It's particularly useful for analyzing complex models like Vision Transformers, enabling the extraction of all attention maps across multiple layers with minimal effort.

Youtu-GraphRAG is a revolutionary framework designed for graph retrieval-augmented complex reasoning, offering a vertically unified agentic paradigm. It jointly connects the entire framework as an intricate integration based on graph schema, allowing seamless domain transfer with minimal intervention. The tool boasts a 33.6% lower token cost and 16.62% higher accuracy over state-of-the-art baselines, making it ideal for multi-hop reasoning, summarization, and knowledge-intensive tasks. Key innovations include schema-guided hierarchical knowledge tree construction, dually-perceived community detection, and agentic retrieval with iterative reflection. It also provides advanced construction and reasoning capabilities for real-world deployment, including user-friendly visualization and parallel sub-question processing.

Qwen-Edit-2509-Upscale-LoRA is an AI tool designed for image editing and upscaling, leveraging LoRA (Low-Rank Adaptation) techniques to enhance image resolution and add high-quality details. Users can upload an image and fine-tune various parameters such as seed, guidance scale, and inference steps to achieve desired visual outcomes. This customization allows for precise control over the enhancement process, making it suitable for individuals looking to improve the quality and detail of their images. The tool aims to provide a flexible solution for image enhancement.

Heuris is an AI-powered learning platform designed to make complex subjects accessible through short, engaging 5-minute sessions. It covers a wide range of topics including History, Art, Economics, and Philosophy. The platform leverages AI conversations to guide users through their learning journey, adapting to individual interests and helping to explore new concepts effectively. This approach makes learning efficient and personalized, catering to individuals who enjoy self-directed exploration and quick, digestible content. Heuris aims to provide an interactive and adaptive educational experience for those looking to expand their knowledge in various academic and cultural fields.

So Vits Svc Models Pcr is an AI tool hosted on Hugging Face Spaces, designed for voice cloning and the creation of custom voice models. While the live website indicates a runtime error and scheduling failure, suggesting current unavailability, the tool's purpose is to enable users to experiment with and develop unique voice models. It is suitable for individuals interested in voice synthesis, research, and development within the AI audio domain. The platform's nature implies a focus on providing a space for community-driven machine learning applications, making it potentially valuable for those looking to explore or contribute to AI voice technology.

awesome-AI-books is a comprehensive GitHub repository dedicated to providing a curated list of AI-related books and PDFs. It serves as an invaluable resource for students and researchers looking to learn and download materials on artificial intelligence. The repository covers a wide range of topics, including introductory AI theory, mathematics for AI, data mining, machine learning, deep learning, philosophy of AI, quantum AI, and various AI frameworks and libraries. It also features a 'Training ground' section with links to platforms for AI experimentation and research, such as OpenAI Gym and DeepMind Pysc2. All books and PDFs are stored on Yandex.Disk due to GitHub's large file storage limitations, and the repository is intended for learning purposes only.

Chinese-Text-Classification-Pytorch is an open-source toolkit designed for Chinese text classification tasks, built on the PyTorch framework. It offers out-of-the-box implementations of several popular text classification models, including TextCNN, TextRNN, FastText, TextRCNN, BiLSTM_Attention, DPCNN, and Transformer. The toolkit is user-friendly and ready for immediate deployment, supporting both character-level input and the integration of pre-trained word vectors, specifically using Sougou News Word+Character 300d. It also includes a pre-processed Chinese dataset (THUCNews) for training and evaluation, making it a comprehensive resource for researchers and developers working on Chinese NLP.

awesome-explainable-graph-reasoning is an open-source collection of research papers and software dedicated to explainability in graph machine learning. This repository serves as a valuable resource for academics and researchers interested in understanding and implementing explainable AI within graph-based models. It categorizes content into explainable predictions, explainable reasoning, software, and theoretical/survey papers, offering a comprehensive overview of the field. The project is licensed under Apache 2.0, making its resources freely accessible for study and development. It's an excellent starting point for anyone looking to delve into the complexities of interpreting graph neural networks and their applications.

The Artificial-Intelligence-Terminology-Database is a comprehensive, open-source mapping database of English to Chinese technical vocabulary in the artificial intelligence domain. Developed by Jiqizhixin, it aims to assist researchers, translators, and students in accurately understanding and translating AI terminology. The database currently contains over 2400 professional terms, with specialized sections for Machine Learning and AI for Science. It provides indexed terms with English and Chinese translations, common abbreviations, and sources/expansions for conceptual understanding. The project emphasizes accuracy, drawing from authoritative textbooks and literature, and encourages community contributions to continuously improve and expand the terminology.