Research & Education

transformers_tasks is an open-source project on GitHub that integrates various NLP algorithms using the powerful Hugging Face transformers library. It offers implementations for a wide range of tasks, including text matching (PointWise, DSSM, Sentence Bert, SimCSE), information extraction (UIE), prompt tasks (PET, p-tuning), and text classification (BERT-CLS). The project also delves into advanced areas like Reinforcement Learning from Human Feedback (RLHF) for language models, text generation (T5-Based models), and large language model (LLM) applications and training. It provides a flexible framework for researchers and developers to train and fine-tune models using their own datasets.

XrayGPT is an advanced AI tool designed for the automated summarization of chest radiographs. It leverages medical vision-language models to process and interpret X-ray images, generating interactive and concise summaries from complex free-text radiology reports. The tool fine-tunes large language models (LLMs), specifically Vicuna, on extensive medical datasets, including 100,000 real patient-doctor conversations and approximately 30,000 radiology conversations. This specialized training allows XrayGPT to acquire domain-specific and relevant features, significantly enhancing the quality and accuracy of its summaries. By aligning a frozen medical visual encoder (MedClip) with a fine-tuned LLM, XrayGPT provides a powerful solution for medical professionals and researchers seeking efficient and accurate analysis of chest X-rays.

VAR (Visual Autoregressive Modeling) is an open-source project that introduces a novel approach to image generation, moving beyond traditional raster-scan "next-token prediction" to a coarse-to-fine "next-scale prediction." This method allows GPT-style autoregressive models to achieve state-of-the-art results, even outperforming diffusion models in visual generation. The project emphasizes scalability, user-friendliness, and provides a robust codebase for researchers and developers. It also highlights the discovery of power-law scaling laws within VAR transformers and demonstrates strong zero-shot generalizability. VAR has received the NeurIPS 2024 Best Paper Award and offers various pre-trained models for different resolutions and complexities.

LunarLink AI provides a unified platform to access and compare outputs from various advanced AI models, including ChatGPT, Claude, and Gemini. It operates on a pay-as-you-go model, charging based on usage at first-party API prices plus a small per-answer fee, eliminating the need for subscriptions or commitments. Users can chat with multiple AI assistants simultaneously, view responses side-by-side to reduce bias, and enjoy features like unlimited file uploads and cross-platform chat synchronization. The platform prioritizes privacy with a data-safe mode that ensures no storage or training of user data, and offers a customizable interface with dark/light modes and enhanced content presentation for rich text and code blocks.

Project December is an AI tool designed to simulate conversations with various individuals, leveraging advanced AI technology. The platform offers text-based interactions, allowing users to engage with AI-powered personas. Its original version, Project December Classic, specifically focuses on simulating conversations with individuals from different eras, providing a unique historical interaction experience. The tool utilizes super-computers to ensure realistic and engaging simulations, aiming to create a compelling conversational experience for its users. It provides an API for integration and offers samples of its conversational capabilities.

STREEKX is an AI search engine designed to provide unrestricted access to information. This tool stands out by offering unlimited usage and search capabilities, allowing users to explore and query information without any imposed limitations. It aims to empower users with a seamless and unhindered search experience, making it suitable for those who require extensive and continuous access to information for research, learning, or general inquiry. The platform focuses on delivering a straightforward and accessible search solution, emphasizing freedom from typical usage constraints often found in similar AI-powered tools.

AI_Books is a GitHub repository offering a comprehensive collection of books focused on Artificial Intelligence, Machine Learning, Deep Learning, and Neural Networks. This resource is designed for individuals looking to deepen their understanding and skills in these rapidly evolving technological domains. The repository includes a wide array of titles, covering foundational concepts, advanced theories, and practical applications, often with examples in popular programming languages like Python and R. It also provides links to online books and supplementary resources, making it a central hub for learning and reference in AI-related subjects.

BeatAI is an open-source educational resource designed to demystify artificial intelligence for a broad audience, from students to engineers. It covers a comprehensive range of topics, including neural networks, large language models, top-level design principles, microscopic mechanisms, engineering implementation, and algorithmic foundations. The project aims to help users understand the profound impact of next-token prediction and realize that AI is not as mysterious or unattainable as it might seem. The resource is presented as an AI entry-level bible, making complex concepts accessible without being overly obscure or niche. It is hosted on GitHub and encourages community contributions.

Awesome-Diffusion-Model-Based-Image-Editing-Methods is an open-source repository offering an exhaustive overview of existing methods using diffusion models for image editing. It covers both theoretical and practical aspects, categorizing works by learning strategies, user-input conditions, and specific editing tasks. The resource pays special attention to image inpainting and outpainting, analyzing traditional context-driven and current multimodal conditional methods. It also introduces EditEval, a systematic benchmark with an innovative LMM Score, to evaluate text-guided image editing algorithms. The repository actively tracks the latest research and welcomes contributions, serving as a valuable reference for researchers and developers in the field.

chatgpt-corpus offers a comprehensive Chinese corpus designed for training large language models. This open-source resource includes diverse datasets such as dialogue, novel, and customer service conversations, totaling millions of entries. The corpus is generated using ChatGPT3.5, providing high-quality data for researchers and developers. It aims to enhance the performance of AI models in various Chinese language tasks, making it a valuable asset for anyone working on natural language processing in Chinese. The project also provides access to related resources and community support.

Deeplearning.ai-Natural-Language-Processing-Specialization is an open-source repository containing comprehensive notes and completed assignments for Coursera's Natural Language Processing Specialization, offered by deeplearning.ai. Taught by instructors Younes Bensouda Mourri and Łukasz Kaiser, the specialization covers four key courses: Natural Language Processing with Classification and Vector Spaces, Probabilistic Models, Sequence Models, and Attention Models. This resource is ideal for students and learners looking to deepen their understanding of NLP concepts, from sentiment analysis and word embeddings to advanced machine translation and chatbot development using models like T5, BERT, and Transformer. It provides practical examples and code structures to aid in learning and application.

Lumina-mGPT 2.0 is an open-source, stand-alone, decoder-only autoregressive model designed for a broad spectrum of image generation tasks. Trained from scratch, it supports functionalities such as text-to-image generation, image pair generation, subject-driven generation, multi-turn image editing, controllable generation, and dense prediction. The project provides inference code for image-to-image tasks and all-in-one model checkpoints on HuggingFace. It also offers acceleration strategies like Speculative Jacobi Decoding and Model Quantization to optimize inference time and GPU memory usage. Lumina-mGPT 2.0 is ideal for AI researchers and machine learning engineers looking to explore and implement advanced image modeling techniques.

Luotuo-Chinese-LLM is an open-source initiative focused on advancing Chinese large language models. The project, developed by researchers from Huazhong Normal University and SenseTime, encompasses a range of models, datasets, pipelines, and applications. Key sub-projects include ChatHaruhi for character-based conversational AI, Luotuo Embedding for generative text embedding, Luotuo QA for conversational question answering, and Mini Luotuo for distilled instruction-following models. It also features Silk Road for building Chinese LLM data foundations and Silk Magic Book for collecting effective prompts. The project emphasizes practical applications and research into cross-language data tuning.

ICLR2025-Papers-with-Code is a comprehensive GitHub repository dedicated to compiling research papers and their corresponding open-source projects from the International Conference on Learning Representations (ICLR). The collection spans from ICLR 2021 to the upcoming ICLR 2025, with a particular emphasis on advancements in Large Language Models (LLMs) and various subfields within Natural Language Processing (NLP). This resource serves as a valuable hub for researchers, academics, and developers looking to stay updated on the latest research trends and access practical code implementations. The repository is actively maintained and updated, encouraging community contributions through watching, forking, and starring the project.

iir is an open-source project hosted on GitHub, offering a collection of algorithms and functionalities for machine learning, natural language processing, and information retrieval. Developed primarily in Python, Ruby, C++, and R, it serves as a valuable resource for researchers and developers in AI-related fields. The repository includes implementations for tasks such as active learning, clustering, natural language detection, LDA, PCA, perceptron, and various neural network components. Its modular structure allows users to explore and integrate different techniques for their specific AI projects, making it suitable for both academic research and practical application development.

The McGill Artificial Intelligence Society (MAIS) is a student-run organization dedicated to making artificial intelligence more accessible to students. They achieve this by hosting a variety of initiatives, including bootcamps like MAIS 202 for ML fundamentals, workshops on applied machine learning topics, and Canada's largest AI hackathon, MAIS Hacks. MAIS also fosters community through events like the Learnathon, an undergraduate AI research conference, and the McGill AI Podcast, which connects ML principles to research disciplines. The society aims to connect McGill students with the broader Montreal AI ecosystem through industry events and networking opportunities.

The National Centre of Artificial Intelligence (NCAI) at UET Lahore, established as the Al-Khwarizmi Institute of Computer Science (KICS) in August 2002, is dedicated to advancing research and development in Computer Science and Information Technology. KICS engages in various research activities, including the development of AI and machine learning solutions. The institute also emphasizes technology transfer through research labs, technology centers, and incubated startups. It actively participates in research collaborations and hosts conferences and workshops, such as the IEEE International Conference on Open Source System and Technologies (ICOSST), to disseminate knowledge and foster innovation in the AI domain.

BELLE, which stands for "Be Everyone's Large Language model Engine," is an open-source initiative by LianjiaTech focused on advancing Chinese dialogue large language models. Unlike projects primarily concerned with pre-training, BELLE emphasizes enabling individuals to create their own high-performing, instruction-following language models based on existing open-source pre-trained models. The project continuously releases instruction training data, relevant models, training code, and application scenarios. It also evaluates the impact of different training data and algorithms on model performance, with a specific optimization for Chinese language using ChatGPT-generated data. Recent updates include enhanced Chinese speech recognition models, multimodal large language models, and research reports on fine-tuning strategies and RLHF training.

Voaige is developing a Test Time Cognition layer for Large Language Models (LLMs), aiming to enhance their reasoning capabilities beyond traditional reinforcement learning and fixed next-token predictions. This innovative approach involves dynamically allocating computational resources during inference, allowing LLMs to perform efficient search and adaptation at test time, similar to how biological cognition navigates complex problems. By understanding and implementing principles from neuroscience, Voaige seeks to enable LLMs to assess difficulty, allocate compute where uncertainty is high, and scale back where it's not, leading to better generalization and the ability to handle novel planning and open-ended complexity without extensive retraining. Their research focuses on architecturally grounded inference systems inspired by the brain's adaptive search mechanisms.

Deep Learning IndabaX South Africa is an annual, locally-organized conference dedicated to fostering knowledge and capacity in machine learning across the African continent. The multi-day event features over 400 attendees, 50+ speakers showcasing cutting-edge research and applied AI, and 4+ tutorials providing practical introductions to machine learning for beginners. Participants can also present their work through 50+ posters, engage in hackathon problems, and connect with a diverse community of researchers and practitioners. The conference aims to make machine learning accessible, offering free attendance options for students and travel grants, supported by various partners and paid registration for academics and industry professionals.

GNNs-for-NLP is a comprehensive resource offering code examples and tutorial materials for applying Graph Neural Networks (GNNs) to Natural Language Processing (NLP) tasks. Originating from presentations at EMNLP 2019 and CODS-COMAD 2020, this GitHub repository provides practical implementations using PyTorch 1.x and TensorFlow 1.x, compatible with Python 3.x. It features simplified GCN model implementations, extensions for relation extraction and word embeddings, and additional resources like theoretical write-ups and recent GNN papers. This tool is ideal for researchers and students looking to understand and implement graph-based deep learning methods in NLP.

LLMBook-zh.github.io is an open-source project providing a comprehensive Chinese textbook titled "大语言模型" (Large Language Models). Authored by Zhao Xin, Li Junyi, Zhou Kun, Tang Tianyi, and Wen Jirong, this book aims to popularize and disseminate the latest advancements in large model technology. It offers a systematic framework and roadmap for LLM technologies, making it ideal for beginners with a background in deep learning. The project includes PDF courseware, code snippets on GitHub, and related teaching videos on Bilibili, serving as a valuable resource for both academic study and practical application. The content covers essential topics such as pre-training, fine-tuning, alignment, and prompt engineering, drawing from the authors' extensive research and practical experience in developing large models.

LLMs_interview_notes is an open-source repository designed to help algorithm engineers prepare for interviews related to Large Language Models (LLMs). The resource compiles learning notes and materials based on personal interview experiences, covering a wide array of LLM topics. It includes detailed sections on LLM fundamentals, advanced attention mechanisms, transformer operations, loss functions, similarity functions, generative LLMs, fine-tuning strategies, LangChain, and Retrieval-Augmented Generation (RAG). This comprehensive collection aims to provide a structured approach to understanding and answering common interview questions in the rapidly evolving field of LLMs.

MLTO: Machine Learning Toronto is a vibrant community dedicated to AI and machine learning enthusiasts within the Greater Toronto Area. It serves as a central hub for professionals, researchers, and individuals passionate about artificial intelligence to connect, share knowledge, and stay updated on the latest advancements. The community regularly hosts events, fostering an environment for networking and collaborative learning. MLTO is committed to building a strong local ecosystem for machine learning, supported by its members and sponsors, and provides guidelines and a code of conduct to ensure a positive and inclusive experience for all participants.