Research & Education

The Arabic RAG Leaderboard, hosted on Hugging Face Spaces, provides a comprehensive platform for evaluating and comparing Arabic Retrieval-Augmented Generation (RAG) systems. This tool is essential for researchers and developers working with Arabic natural language processing, offering insights into how various models perform on critical tasks like information retrieval and re-ranking. Users can easily switch between tabs to analyze the performance metrics of different RAG models, helping them identify the most effective solutions for their specific needs. The leaderboard supports the evaluation of 'No, Full & Late Interaction Models,' providing a nuanced view of model capabilities and limitations in the Arabic language context.

Tune-A-Video Inference is an AI-powered tool hosted on Hugging Face Spaces, designed for generating videos from textual descriptions. Users can input a text prompt and then customize various parameters, including the choice of AI model, desired video length, and frames per second (FPS). This flexibility enables users to experiment with different settings to achieve their desired video output. The platform is particularly useful for AI researchers, developers, and video creators who are interested in exploring and leveraging AI models for video content creation. It provides a straightforward interface for generating unique video content based on textual input, making advanced video generation accessible.

VibeVoice-Realtime-0.5B is an AI-powered tool hosted on Hugging Face that specializes in real-time text-to-speech conversion. Users can input English text and select a speaker voice to generate spoken audio. A key feature is the ability to fine-tune the voice fidelity using a slider, allowing for customization of the output quality. The application provides the generated audio as a downloadable WAV file, making it suitable for various applications requiring spoken content. This tool is designed for quick and efficient audio generation from text.

Visionbotix is a technology company specializing in automation, intelligence, and software development. They offer a range of services including robotics, computer vision, artificial intelligence, and embedded systems. Their expertise extends to developing web, Android, and iOS applications, as well as game development. Visionbotix focuses on creating industry-standard, competitive solutions using cutting-edge technologies, working closely with clients from idea generation to launch. They aim to solve real-world problems by providing smart and automated solutions, such as their livestock management system and custom surveillance monitoring powered by AI-trained cameras.

NLP-Knowledge-Graph is an open-source GitHub repository dedicated to the research and application of natural language processing, knowledge graphs, dialogue systems, and large language models. It serves as a comprehensive resource, offering deep learning insights for knowledge graphs, research summaries, and a curated list of relevant papers. The repository includes practical applications such as building knowledge-graph-based dialogue systems and provides links to various NLP tools, datasets, and visualization utilities. It also covers topics like Chinese financial document processing, event knowledge graphs, and the commercialization of NLP/dialogue/KG technologies, making it a valuable asset for researchers and developers in the field.

Vevo is an AI-powered tool hosted on Hugging Face Spaces, designed for controllable zero-shot voice imitation. It enables users to transform the style and timbre of an audio file by providing a reference audio file. This functionality is useful for voice cloning and text-to-speech applications, allowing for a high degree of control over the output audio. The tool requires users to upload two audio files: one for the content and another for the desired style or timbre. While the platform experienced a runtime error at the time of scraping, its core offering focuses on advanced audio manipulation for creative and practical purposes.

VibeVoice ASR is an official playground for Microsoft's VibeVoice-ASR, an advanced AI tool designed for automatic speech recognition. Hosted on Hugging Face Spaces, this application enables users to easily convert spoken language into written text. Users can input either pre-recorded audio files or utilize live speech, and the system will generate precise text transcriptions. This tool is ideal for anyone needing to quickly and accurately transcribe audio, making it a valuable resource for various applications ranging from content creation to documentation.

WebAssembly English TTS (sherpa-onnx) is a text-to-speech tool hosted on Hugging Face Spaces that allows users to convert English text into spoken audio. The unique aspect of this tool is that it runs the speech-synthesis model entirely locally within your browser using WebAssembly. This means all processing happens on your device, ensuring privacy and instant audio generation. Users can type the desired text, adjust parameters like speaker ID and speech speed, and then generate an audio clip that can be played immediately. It's an efficient solution for generating speech without relying on external servers for processing.

WhyBot is an AI-powered assistant designed to facilitate in-depth exploration of various questions and topics. Users can delve into subjects by customizing settings such as the AI's persona and the underlying model, allowing for a tailored conversational experience. The tool aims to provide thoughtful and comprehensive responses, enhancing user understanding and stimulating intellectual curiosity. It is particularly beneficial for individuals engaged in research, learning, or anyone with a curious mind seeking to explore subjects from different perspectives.

cgft-llm is an open-source GitHub repository dedicated to providing practical learning resources for large language models (LLMs). It offers a comprehensive series of code examples, detailed documentation, and video tutorials covering various aspects of LLM development, including Agent systems, core LLM technologies like fine-tuning and RAG, and integration with open-source projects. The resource is structured into several modules, guiding users through topics such as LLM deployment with tools like vLLM and Ollama, data preparation for fine-tuning, and advanced concepts like function-calling and tool-use. It also includes sections on Python engineering practices for open-source contributions and creative AI projects, making it suitable for developers and researchers looking to deepen their understanding and practical skills in the LLM domain.

Linly is an open-source project that offers a suite of Chinese large language models, including Chinese-LLaMA (versions 1 and 2), Chinese-Falcon, and Linly-OpenLLaMA. These models are built upon LLaMA and Falcon architectures, enhanced with extensive Chinese and English parallel corpora for incremental pre-training, extending their linguistic capabilities to Chinese. The project also features Linly-ChatFlow, a Chinese conversational model trained on large-scale instruction data. Linly provides comprehensive resources, including data preparation, model training, and evaluation code, ensuring reproducibility. It supports various quantization schemes for deployment on CUDA and edge devices, making it versatile for different applications. The Linly-OpenLLaMA models, available in 3B, 7B, and 13B scales, are trained from scratch on 1TB of mixed English and Chinese data, featuring an optimized tokenizer for Chinese characters and words, and are released under an Apache 2.0 license for commercial use.

MemMachine Playground is an official platform by Memverge designed for experimenting with AI memory for large language models (LLMs) and AI agents. This tool allows users to explore various memory configurations and understand their impact on AI performance and capabilities. It serves as a valuable resource for AI developers and researchers who are focused on enhancing the intelligence and efficiency of their AI systems. The playground offers a hands-on environment to test and refine memory strategies, ultimately contributing to the development of more robust and effective AI applications.

PaddleNLP is a comprehensive development suite built on the PaddlePaddle deep learning framework, designed for large language models (LLMs). It facilitates efficient training, lossless compression, and high-performance inference of models across diverse hardware, including NVIDIA GPUs, Kunlun XPU, Ascend NPU, and more. The library emphasizes ease of use and extreme performance, aiming to empower developers in creating industrial-grade large model applications. Key features include 4D high-performance training with data, tensor, and pipeline parallelism, efficient fine-tuning algorithms, and a high-performance inference module with dynamic insertion and operator fusion. It also supports a wide range of popular LLM series like LLaMA, Baichuan, Bloom, ChatGLM, Gemma, Mistral, OPT, and Qwen.

ML-NLP is an open-source GitHub project designed to be a comprehensive resource for individuals preparing for interviews in Machine Learning, Deep Learning, and Natural Language Processing. It covers frequently tested knowledge points and provides practical code implementations, making it an invaluable theoretical foundation for aspiring algorithm engineers. The project is structured into various modules, offering a clear and organized knowledge system. Each chapter includes potential interview questions and concludes with practical algorithm case studies. It is intended for continuous learning, review, and as a quick reference during interview preparation.

Zhihu is a GitHub repository offering a collection of source code for various Artificial Intelligence projects, primarily focusing on Natural Language Processing (NLP) and Computer Vision (CV). Implemented in Python 3.6, this repository serves as a practical resource for developers and researchers. It features diverse projects such as text generation using RNNs (LSTM), machine translation with Seq2Seq models and attention mechanisms, deep convolutional Generative Adversarial Networks (GANs) for image generation, and sentiment analysis using DNN, LSTM, and CNN. Additionally, it includes implementations for image style transfer and CTR prediction models like DeepFM and xDeepFM. The repository is designed to accompany a personal column, providing hands-on code examples for learning and experimentation in AI.

Videmak Research AI is an advanced AI-powered platform designed to significantly shorten the academic research journey. It provides a comprehensive suite of tools for students and researchers, enabling them to generate research proposals, conduct literature reviews, create citations, and perform data analysis in minutes rather than hours. The platform integrates AI Chat Bots trained on academic expertise to provide instant answers and information. Beyond core research functions, Videmak also offers content generation tools like article writers, rewriters, and summarizers, alongside project and team management features, making it a versatile solution for academic endeavors.

MLGym is an experimental framework and benchmark designed for advancing AI Research Agents, particularly focusing on reinforcement learning (RL) algorithms for training such agents. It provides the first Gym environment specifically tailored for machine learning tasks. The platform features MLGym-Bench, a collection of 13 diverse and open-ended AI research tasks spanning domains like computer vision, natural language processing, reinforcement learning, and game theory. These tasks are designed to challenge agents with real-world AI research skills, including idea generation, data processing, ML method implementation, model training, experimentation, and iterative improvement. Currently under heavy development by GenAI at Meta and UCSB NLP, MLGym aims to expand the selection of AI research tasks for benchmarking LLM Agents and implementing RL algorithms in a research environment. It supports containerized execution via Docker or Podman and offers a Web UI for trajectory visualization.

ALCE, pronounced /elk/, is an open-source project from Princeton NLP designed to enable large language models (LLMs) to generate text with accurate citations. It introduces a benchmark for Automatic LLMs' Citation Evaluation (ALCE) and includes three datasets: ASQA, QAMPARI, and ELI5. The repository provides comprehensive code for automatic evaluation of LLM generations across fluency, correctness, and citation quality. Researchers can also reproduce baselines from the associated EMNLP 2023 paper, perform passage retrieval, and add post-hoc citations to closed-book models. It supports both OpenAI API and offline HuggingFace models, making it a versatile tool for academic research in natural language processing.

MT-DNN (Multi-Task Deep Neural Networks) is a PyTorch-based package designed for Natural Language Understanding. It implements advanced techniques such as adversarial training for both LM pre-training/finetuning and f-divergence, as well as large-scale adversarial training for language models (ALUM). The tool also features SMART, a principled regularized optimization method for robust and efficient fine-tuning of pre-trained natural language models. MT-DNN supports various NLP tasks including GLUE benchmark reproduction, SciTail & SNLI domain adaptation, sequence labeling, and question answering. It provides functionalities for extracting text embeddings and offers options for speeding up training through gradient accumulation and FP16 support.

nlp-notebooks is an Open Source collection of Jupyter notebooks designed for Natural Language Processing (NLP) education and practical application. It offers a comprehensive set of resources from NLP Town, covering fundamental concepts like word embeddings and advanced topics such as named entity recognition with various models including BiLSTM and Transformers. The collection also delves into text classification using traditional methods like Scikit-learn, as well as modern approaches with BERT and CNNs. Users can explore multilingual embeddings, cross-lingual sentence similarity, and transfer learning. It's an invaluable resource for students and practitioners looking to understand and implement NLP techniques.

llms-interview-questions is an open-source resource offering a curated collection of interview questions and detailed answers specifically focused on Large Language Models (LLMs). Designed to assist job seekers in the machine learning and data science fields, this repository covers fundamental LLM concepts, architectures like Transformers (including modern 2026 standards), training pipelines (pre-training, SFT, DPO/RLHF), and key differences from traditional statistical language models. It delves into technical specifics such as RMSNorm, SwiGLU activation functions, Rotary Positional Embeddings (RoPE), and Grouped-Query Attention (GQA), providing code examples for better understanding. The resource aims to equip candidates with the knowledge needed to excel in technical interviews.

The UCLA Smart Grid Energy Research Center (SMERC) is a leading academic research institution dedicated to advancing knowledge in artificial intelligence and machine learning applications for the energy and transportation sectors. SMERC's research encompasses critical areas such as electric vehicle (EV) fleet charging, vehicle-to-grid (V2G) technologies, seamless integration of solar and battery storage systems, and the development of resilient microgrids. Additionally, the center explores the role of AI in autonomous EVs. Through its innovative research, SMERC aims to foster a sustainable future for both transportation and energy systems, addressing pressing societal needs and contributing to global engagement in these vital fields.

MGM Omni is a Hugging Face Space designed to scale Omni LLMs for personalized, long-horizon speech generation. This application enables users to create voice responses that accurately match a provided reference voice. Users can either input text directly or upload existing audio to generate the desired personalized speech. The tool supports bot integration, making it suitable for various applications requiring custom voice output. It is intended for research and development in speech technology, offering a platform to explore advanced voice synthesis and personalization.

Multilingual LLM Tokenizers is an AI development tool designed for AI developers and NLP researchers to experiment with and understand tokenization processes. Users can enter or upload multilingual text to see its tokenized form and get detailed statistics. The tool provides insights into total tokens, token types, and compression ratio, which are crucial for optimizing language models and understanding their behavior across different languages. It supports research and development in multilingual natural language processing, offering a practical way to visualize and analyze tokenizer performance.