Research & Education

CL EVA02 LoRA ONNX Tagger is an AI tool designed for image tagging, specifically for anime images and illustrations. Users can upload an image or provide an image URL to receive predicted tags that describe its content. The tags are categorized into types such as rating, general, and character. The tool also offers a visualization of the generated tags, providing a comprehensive overview of the image's characteristics. It utilizes ONNX models for efficient image classification, making it suitable for tasks like organizing image datasets and supporting computer vision research.

Dimple 7B is a discrete diffusion multimodal large language model designed for image-text-to-text tasks. This application enables users to upload images and type questions or prompts, receiving informative answers and detailed responses. Built upon Dream-org/Dream-v0-Instruct-7B, Dimple 7B has been trained on extensive datasets such as LLaVA-CC3M-Pretrain-595K and Lmms-lab/LLaVA-NeXT-Data, ensuring robust performance in multimodal understanding and generation. It provides a platform for advanced AI interactions, bridging the gap between visual and textual information to deliver comprehensive outputs.

ReasonFlux is an advanced open-source LLM post-training suite developed by a collaboration of Princeton University, PKU, UIUC, University of Chicago, and ByteDance Seed. Its core mission is to build next-generation reasoning capabilities by focusing on innovative algorithms for data selection, reinforcement learning, and inference scaling. The suite includes ReasonFlux-PRM, which offers trajectory-aware process reward models for long Chain-of-Thought (CoT) reasoning, providing dense supervision for data selection and policy optimization. ReasonFlux-Coder introduces a co-evolutionary reinforcement learning approach for LLM coders and unit testers, leading to more robust coding capabilities. Additionally, the suite incorporates preliminary work on thought templates, such as Buffer of Thoughts and SuperCorrect, to guide complex problem-solving and achieve state-of-the-art performance with higher efficiency.

text-classification-surveys is an open-source GitHub repository dedicated to compiling extensive resources for text classification within Natural Language Processing (NLP). It offers a detailed overview of various models, ranging from deep learning approaches like SpanBERT, ALBERT, and BERT, to shallow learning techniques such as LightGBM, SVM, and Random Forest. The repository also covers a wide array of text classification datasets, including MR, SST, IMDB, and Yelp, alongside common evaluation metrics like accuracy, Precision, Recall, and F1. Furthermore, it addresses technical challenges, including multi-label text classification. The content is primarily derived from the paper "A Survey on Text Classification: From Shallow to Deep Learning," making it a valuable resource for researchers and students in the field.

Talk To Qwen Webrtc is an AI tool designed for real-time voice interaction with the Qwen2Audio model, leveraging Gradio and WebRTC technologies. Users can speak into a microphone, and the application will transcribe their speech into text. Following transcription, the tool processes the audio input and generates a text-based response, enabling dynamic communication with an AI. This platform is hosted on Hugging Face Spaces, making it accessible for experimentation with AI-driven audio processing and voice agents. It offers a straightforward interface for those looking to explore speech-to-text and AI response generation capabilities.

Reinforcement-Learning-Papers is an open-source GitHub repository that serves as a curated collection of research papers in the field of reinforcement learning. It encompasses both foundational classic papers and the latest research presented at top conferences such as ICLR, ICML, and NeurIPS. The collection primarily focuses on single-agent reinforcement learning, offering a structured overview of various topics including Model-Free (Online) RL, Model-Based (Online) RL, Offline RL, Meta RL, Adversarial RL, and RL with Transformer/LLM. This resource is invaluable for researchers, academics, and students who need to stay updated on significant advancements and foundational concepts in reinforcement learning.

scGPT is an open-source codebase designed to build a foundation model for single-cell multi-omics data using generative AI. It offers pre-trained models for various human cell types and organs, including whole-human, brain, blood, heart, lung, kidney, and pan-cancer. The tool supports zero-shot applications for cell embedding tasks and features efficient reference mapping for millions of cells. Researchers can install scGPT via pip and utilize online apps for reference mapping, cell annotation, and Gene Regulatory Network inference. It is ideal for bioinformaticians and computational biologists working with complex single-cell datasets.

Emu2 is a generative multimodal model developed by BAAI, designed for in-context learning and capable of processing both image and text inputs. This application, hosted on Hugging Face Spaces, enables users to generate various forms of content and engage in interactive chat experiences. By providing a combination of text and images, users can receive generated responses or participate in conversations, making it a versatile tool for multimodal AI research and experimentation. The model aims to push the boundaries of AI's ability to understand and create content across different modalities.

rwa is an open-source recurrent neural network (RNN) model designed for machine learning on sequential data. It introduces a novel approach by computing a recurrent weighted average (RWA) over every previous processing step, allowing for direct connections anywhere along a sequence. This method contrasts with traditional RNN architectures that typically rely only on the immediate previous step. The RWA model can be computed as a running average, meaning it doesn't need complete recomputation at each step, leading to scalability comparable to LSTM models. Notably, rwa demonstrates considerably faster training on most tasks, often by a factor of five or more, with performance improvements scaling further for longer sequences. While effective for many tasks, it has not yielded competitive results for Natural Language Problems.

The HKUST(GZ) Information Hub, part of The Hong Kong University of Science and Technology (Guangzhou), is dedicated to advancing information science and technology through innovative education and research. It emphasizes a cross-disciplinary approach, integrating various fields to address complex challenges. The hub focuses on key areas such as Artificial Intelligence, Data Science, Internet of Things (IoT), and Computational Media, aiming to cultivate forward-looking talents with a global vision. Through its academic structure, which includes various Hubs and Thrust Areas, HKUST(GZ) seeks to promote higher education reform and accelerate collaboration between Hong Kong and Mainland China, contributing significantly to the development of the Greater Bay Area.

Chainer Chemistry is an open-source deep learning library specifically designed for applications in biology and chemistry. Built upon the Chainer framework, it provides support for a variety of state-of-the-art models, with a particular focus on Graph Convolutional Neural Networks (GCNN) for chemical property prediction. The library enables researchers and developers to train new models on given datasets, perform inference using pre-trained models, and evaluate the performance of different models. It supports numerous GCNN architectures like NFP, GGNN, WeaveNet, SchNet, and MPNN, along with several chemical and network datasets such as QM9, Tox21, MoleculeNet, and ZINC. Chainer Chemistry is a valuable tool for academic research, facilitating advancements in areas like molecular graph generation and the explanation of GCNN predictions.

Groq-LLaMA3/4 is a chat application built using Streamlit and the Groq API, hosted as a Hugging Face Space. It enables users to engage with various Llama models and other AI models available through a Groq account. The tool supports multimodal interactions, allowing users to upload and reference .txt, .md, or .pdf files within their conversations. Additionally, if the selected model has vision capabilities, users can upload images for analysis. This makes it a versatile platform for exploring and interacting with advanced AI models in a conversational format.

This GitHub repository, self-correction-llm-papers, serves as a comprehensive collection of research papers focused on the self-correction mechanisms of Large Language Models (LLMs) using automated feedback. It is an invaluable resource for researchers and practitioners delving into the advancements and methodologies for improving LLM performance and reliability. The collection is meticulously organized into categories such as Training-Time Correction (including RLHF, Fine-tuning, and Self-Training strategies), Generation-Time Correction (covering Re-Ranking and Feedback-guided strategies), and Post-hoc Correction (featuring Self-Refine, External Feedback, and Model-Debate strategies). This structured approach allows users to easily navigate and explore the landscape of diverse self-correction techniques.

Giant Music Transformer is a powerful AI tool designed for generating multi-instrumental music. Users can initiate music creation by uploading an existing MIDI file or by starting with a random input, offering flexibility in the creative process. The tool provides various customizable settings, including the number of tokens, temperature, and drum introduction, allowing for fine-tuned control over the generated output. This makes it suitable for musicians, content creators, and anyone looking to experiment with AI-driven music composition. The application is hosted on Hugging Face Spaces and is available under the Apache 2.0 license, promoting open access and collaboration.

Gemma-3-R1984-27B ChatBot is an AI-powered application designed to provide answers by analyzing various document types, including text, PDF, CSV, and TXT files. Users can upload their documents and then ask questions, receiving detailed responses derived directly from the content. This tool is built for reasoning and deep research, leveraging the Gemma-3 family of models. It is hosted on Hugging Face Spaces and benefits from the processing power of NVIDIA H100 GPUs, indicating a focus on robust performance for complex analytical tasks. The application aims to streamline information extraction and question-answering from diverse data sources.

The QUT Urban AI Hub is a dedicated research center within the Queensland University of Technology, focusing on the intersection of artificial intelligence and urban development. Its primary goal is to foster research into smart and sustainable cities and communities. The hub serves as a central point for researchers exploring various topics related to urbanisation and AI, offering a supportive network and resources. It aims to drive innovation and provide solutions for real-world urban challenges by leveraging AI technologies and interdisciplinary collaboration within the academic community.

Sentiment-Analysis-Twitter is an open-source research project focused on determining the best combination of feature sets and machine learning classifiers for sentiment analysis of Twitter data. The project explores various pre-processing steps, including handling punctuations, emoticons, Twitter-specific terms, and stemming. It investigates features such as unigrams, bigrams, trigrams, and negation detection, and trains classifiers using algorithms like Naive Bayes, Decision Trees, and Maximum Entropy. The methodology uses a modularized approach, separating feature extraction and classification, which allows for flexible experimentation. While the original project code remains publicly hosted, the project has been sold and is also available as a commercial API through OnePanel Inc.

sentiment_analysis_fine_grain is an open-source GitHub repository providing resources for multi-label classification with BERT, specifically tailored for fine-grained sentiment analysis. It includes code for training models, deploying BERT for online prediction, and a tutorial for using BERT with Chinese text. The repository offers pre-trained models, data for fine-tuning and pre-training, and scripts for data preprocessing and model evaluation. It supports both BERT and TextCNN architectures, allowing users to experiment with different approaches to sentiment analysis. This tool is ideal for researchers and NLP engineers working on sentiment analysis tasks, particularly those involving Chinese language data.

Reach Industries focuses on building frontier technologies and designing systems that prioritize human collaboration. Their initial venture is in the science sector, where they've developed Lumi, the first Visual AI Copilot for science. Lumi aims to upgrade scientific processes by addressing the complexities and heavy regulations within laboratories, which often rely on manual record-keeping and human observation for critical tasks. By integrating AI, Lumi seeks to improve efficiency and accuracy in scientific research, allowing scientists to concentrate on higher-level work. The company's vision is to foster a future where humans and machines work together seamlessly, enhancing industries and unlocking human potential.

StableNormal is an open-source AI tool designed to enhance monocular normal estimation by reducing the inherent stochasticity of diffusion models. This approach leads to "Stable-and-Sharp" normal maps, outperforming various baselines in terms of accuracy and stability. The tool is presented as a research project from SIGGRAPH Asia 2024 and provides a Python-based pipeline for installation and usage. It includes a faster inference option, StableNormal-turbo, which is 10 times quicker. Users can compute metrics on datasets like DIODE, IBims-1, Scannet, and NYUv2 to evaluate performance, making it suitable for researchers and developers in computer vision and generative AI.

GrantGPT is an AI platform specifically designed to explore European funding opportunities, acting as a public funding matchmaker. It provides an up-to-date database of active calls for proposals, complete with citations, allowing users to research and apply in minutes. The platform can summarize key information from European calls and automatically draft documents, emails, and reports, saving time for analysis. GrantGPT uses AI to find EU funding opportunities, known as calls for proposals, and aims to make expert knowledge about European funding accessible. It is built with high security and privacy standards, is GDPR compliant, and hosted in Germany. The tool offers a basic free search function with additional premium services for a comprehensive funding search experience.

SpikeGPT is an implementation of a generative pre-trained language model that utilizes pure binary, event-driven spiking neural networks. This lightweight model is inspired by RWKV-LM and allows for experimentation with spiking neural networks in language modeling tasks. It supports training on datasets like Enwik8 and pre-training on large corpora such as The Pile. Users can fine-tune the model on datasets like WikiText-103 and perform inference with custom prompts or a pre-trained model. The repository also includes resources for fine-tuning with Natural Language Understanding (NLU) tasks, making it a valuable tool for researchers and developers exploring alternative neural network architectures.

SuperGlue-pytorch offers a PyTorch implementation of the SuperGlue matching network, designed for learning feature matching with Graph Neural Networks. This repository specifically includes code for training the SuperGlue network using SIFT keypoints and descriptors. It is intended for applications leveraging the Physarum Dynamics LP solver, which can potentially replace the original Sinkhorn Algorithm in SuperGlue. The architecture involves an Attentional Graph Neural Network and an Optimal Matching Layer, facilitating the identification of correspondences between image features, even in cases of occlusion or detector failure. The tool provides scripts for training the model and loading data, including generating keypoints, descriptors, and ground truth matches.

symbolic_deep_learning is an open-source project providing the official implementation for the research paper "Discovering Symbolic Models from Deep Learning with Inductive Biases." This tool enables researchers and developers to explore the integration of symbolic reasoning with deep learning techniques. It supports the development of models that combine neural networks with symbolic structures, offering a novel approach to understanding and interpreting complex deep learning models. The repository includes code for training example models, generating data, and analyzing results, making it a valuable resource for academic research in AI and machine learning.