Research & Education

BERT-flow offers a TensorFlow implementation of the research paper "On the Sentence Embeddings from Pre-trained Language Models" (EMNLP 2020). This tool is designed for researchers and developers working with natural language processing, specifically focusing on enhancing the quality of sentence embeddings derived from pre-trained BERT models. It provides scripts and configurations for fine-tuning BERT with NLI supervision and for unsupervised learning of flow-based generative models. The repository includes detailed instructions for setting up the environment, downloading pre-trained BERT models and GLUE data, and running experiments for both fine-tuning and flow-based model training and evaluation. BERT-flow is a valuable resource for academic research and experimentation in the field of sentence representation.

Cell2Sentence (C2S) is an open-source framework designed for applying Large Language Models (LLMs) to single-cell transcriptomics. It implements the C2S-Scale framework, which transforms expression vectors into "cell sentences"—space-separated gene names ordered by descending expression. This innovative approach allows LLMs to natively model scRNA-seq data using natural language, unifying transcriptomic and textual data. The tool enables advanced single-cell tasks such as perturbation prediction, dataset summarization, cluster captioning, and biological question answering. C2S-Scale models, including those based on Pythia and Gemma-2 architectures, are available on Huggingface, with support for finetuning on custom prompt templates and multi-cell prompt formatting.

Building-llm-applications-from-scratch offers an open-sourced course designed for professionals to master the development of Large Language Model (LLM) applications. Unlike many courses that rely on pre-built frameworks, this program delves into the foundational building blocks of retrieval systems, empowering participants to design, build, and deploy custom LLM-powered solutions from scratch. The curriculum covers essential topics such as Transformer Architecture, Retrieval-Augmented Generation (RAG), and open-source LLM deployment. It includes 29 in-depth lessons, 6 real-world projects, interactive live sessions, and direct instructor access, culminating in a certificate upon completion. The course is ideal for those with existing Python and basic machine learning knowledge.

The Bamberg Center for Artificial Intelligence (BaCAI) is a research institution dedicated to advancing open AI research with national and international visibility. Its core mission involves the responsible translation of AI algorithms into practical applications, with a strong emphasis on developing human-centric AI systems. BaCAI fosters interdisciplinary cooperation to achieve its goals, aiming to become a central hub for AI expertise and talent. The center's work contributes to the broader academic landscape by integrating AI research within the Otto-Friedrich-Universität Bamberg's various faculties, including humanities, social sciences, and applied computer science.

clinicalBERT is an open-source repository offering publicly available Clinical BERT embeddings, designed to advance clinical Natural Language Processing (NLP) research. It enables users to leverage pre-trained models like Bio+Clinical BERT and Bio+Discharge Summary BERT, which are finetuned from BioBERT or the cased version of BERT. The tool provides clear instructions for direct integration via the Hugging Face transformers library, simplifying access for researchers and developers. Additionally, it outlines steps to reproduce the pretraining process using MIMIC data and offers examples for downstream tasks such as Med NLI and NER, making it a comprehensive resource for those working with clinical text data.

CLIP_prefix_caption is an open-source image captioning model that provides a novel approach to generating descriptive captions for images. Unlike traditional methods that often require additional supervision like object annotation, this model only needs images and their corresponding captions for training, making it highly adaptable to various datasets. It leverages the powerful CLIP model for generating semantic encodings and fine-tunes a pretrained language model to produce meaningful sentences. The tool boasts significantly faster training times while maintaining state-of-the-art results, even on large datasets like Conceptual Captions. It also offers a variant using a transformer architecture for the mapping network, avoiding GPT-2 fine-tuning, and still achieving comparable performance on the nocaps dataset. The project provides inference notebooks and a GUI for easy visualization and use.

chatgpt-your-files is an open-source, production-ready MVP for securely chatting with your documents, leveraging OpenAI’s GPT models and retrieval augmented generation (RAG). It features an interactive chat interface for engaging with your documentation, along with robust document storage capabilities for securely uploading, storing, and retrieving user-uploaded files. The tool integrates one-click third-party login with 18 authentication providers and offers a flexible REST API for front-end consumption. Crucially, it incorporates row-level security to protect all user data, ensuring a secure environment for document interaction. The project is presented as a workshop, guiding users through setting up storage, processing documents, creating embeddings, and building the chat functionality.

ChatDocs is an AI-powered tool designed for secure, offline interaction with your documents. It allows users to chat with their documents without any data leaving their local system, making it ideal for handling sensitive information. The tool only requires an internet connection for initial installation and downloading AI models. Based on PrivateGPT, ChatDocs offers enhanced features and supports various AI models including GGML/GGUF via CTransformers, Hugging Face Transformers, and GPTQ models. It provides both a web UI and a command-line interface, and is highly configurable through a `chatdocs.yml` file, allowing users to customize embeddings and LLM models. It also supports GPU acceleration for improved performance.

The AI Times is an innovative AI-powered news application designed to keep users informed with factual and unbiased news. It leverages artificial intelligence to generate news content, offering a fresh perspective on current events. The platform aims to counteract bias often found in traditional news sources by presenting information generated by AI. Additionally, The AI Times incorporates humor into its news delivery, making the consumption of news more engaging and enjoyable. This unique blend of AI-driven factual reporting and lighthearted presentation helps users stay up-to-date in an accessible and entertaining manner.

Deep-Learning-Experiments is an open-source GitHub repository designed to help users understand deep learning through a combination of videos, detailed notes, and practical experiments. It offers comprehensive lecture notes covering fundamental deep learning topics such as Supervised Learning, Multilayer Perceptrons (MLP), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Transformers. The repository also includes code implementations for many of these concepts, allowing users to run and experiment with models like Mamba, Autoencoders (AE), Variational Autoencoders (VAE), Generative Adversarial Networks (GAN), and Diffusion Models. Additionally, it provides resources for setting up development environments, including Python, Numpy, PyTorch, and Hugging Face, making it a valuable resource for both theoretical understanding and practical application in deep learning.

SwissNLP is an association dedicated to advancing Natural Language Processing (NLP), Computational Linguistics, and Text Analytics within Switzerland. It serves as a bridge between AI and human language understanding and application, bringing together experts, solution providers, and customers from both industry and academia. The association organizes various events and projects to foster growth in the NLP field and also distributes datasets created through its initiatives. SwissNLP aims to promote innovation and knowledge sharing, offering membership opportunities for updates and collaboration, including a new 'Young Professionals' membership starting in 2026.

Kartoffel-1B-v0.1-Llasa 1b Tts is an AI tool hosted on Hugging Face Spaces, specializing in German zero-shot voice cloning. Users can generate speech from text by providing a reference audio sample, enabling personalized voice synthesis. The application also offers the flexibility to choose from a selection of predefined speakers or opt for a random voice, providing diverse options for audio output. This tool is fine-tuned with Llasa 1b, ensuring high-quality voice generation. The output is an audio file, making it suitable for various applications requiring synthesized German speech.

Wave is a comprehensive AI note taker and meeting transcription application designed to capture, transcribe, and summarize audio from various sources. It supports meetings, phone calls, lectures, and general conversations, making it ideal for professionals and students alike. The tool operates across a wide range of devices including iPhone, Android, Mac, Windows, and Apple Watch, with automatic syncing across all platforms. Wave offers highly accurate transcriptions in 76 languages, with automatic speaker identification and the ability to translate between languages. Users can customize summary formats, add notes and photos during recording, and import existing audio files, YouTube videos, or PDFs. It also integrates with popular meeting platforms like Zoom, Google Meet, and Microsoft Teams, and offers a Developer API for advanced workflows.

deep-learning-pytorch-huggingface is an open-source GitHub repository dedicated to providing comprehensive instructions, examples, and tutorials for individuals looking to get started with deep learning using PyTorch and Hugging Face libraries. It covers a wide range of topics, including fine-tuning large language models like FLAN-T5 and Falcon 180B with advanced techniques such as DeepSpeed ZeRO, LoRA, and Flash Attention. The repository also includes guidance on using transformers and datasets, quantizing open LLMs with optimum and GPTQ, and implementing RLHF with DPO. It's a valuable resource for learning about efficient distributed training with FSDP and Q-LoRA, as well as various inference examples for text generation and other tasks.

DISC-LawLLM is an intelligent legal system powered by large language models (LLMs), developed and open-sourced by Fudan University's Data Intelligence and Social Computing Laboratory (Fudan-DISC). It offers comprehensive legal services, including legal text processing for information extraction and summarization, and legal reasoning capabilities enhanced by legal syllogism theory. The system also features a retrieval-augmented module for improved knowledge adherence, utilizing a vast knowledge base of laws and legal exam questions. DISC-LawLLM provides high-quality training datasets, effective training paradigms, and a robust evaluation framework, with its performance on the Lawbench benchmark ranking second only to GPT-4 among legal LLMs.

dla is an open-source project offering extensive deep learning materials specifically tailored for audio processing. It provides lecture and seminar content covering a wide array of topics, including digital signal processing, automatic speech recognition (ASR), source separation, text-to-speech (TTS), neural audio codecs, and voice biometry. The repository includes practical exercises and project templates, making it suitable for both theoretical learning and hands-on implementation. Originally conducted at the CS Faculty of HSE, the course materials are organized by week, with some lecture recordings available in English. It serves as a valuable educational resource for students and researchers interested in the application of deep learning to audio.

DL4Proteins-notebooks offers a comprehensive series of Colab notebooks designed to democratize deep learning for protein design and prediction. This resource provides an accessible, hands-on introduction to the tools and methodologies that have revolutionized computational protein science, including AlphaFold, RFDiffusion, and ProteinMPNN. By blending foundational machine learning principles with state-of-the-art approaches, DL4Proteins equips researchers, educators, and students with the knowledge to contribute to the future of protein engineering. The open-source notebooks bridge the gap between cutting-edge research and classroom learning, fostering innovation in synthetic biology and therapeutics. It covers topics from neural networks with NumPy and PyTorch to graph neural networks and diffusion models.

ActionSync is an enterprise AI solution designed to integrate seamlessly with an organization's existing tools like Slack, Gmail, Drive, Notion, Jira, and HubSpot. It acts as an invisible intelligence layer, unifying data, understanding business context, and executing tasks securely and privately. The platform offers specialized AI assistants for various roles, including Sales, Marketing, Product Management, and Engineering, automating routine tasks such as report generation, email drafting, and CRM updates. A key differentiator is its enterprise-grade privacy, running on a privately hosted architecture where data remains within the user's environment, ensuring security and compliance. ActionSync aims to boost operational efficiency, accelerate innovation, and enhance decision-making by turning scattered information into actionable intelligence.

DuckDuckGPT enhances the DuckDuckGo search experience by integrating AI chat and search summaries, powered by the latest large language models. This tool provides users with succinct and comprehensive answers directly within their search results, streamlining the research process. It's designed to save time and boost productivity by delivering precise information faster, making it valuable for various applications from general research to trend analysis. The tool supports multiple browsers and userscript managers, offering flexibility in its installation and use. It also features an optional Proxy API Mode for text responses, removing the need for a ChatGPT account.

Imagine Stories is an innovative AI-powered platform designed for quickly creating personalized texts, stories, and therapeutic fairy tales for children aged 3 to 16. It caters to teachers, therapists, parents, psychologists, pedagogues, and speech therapists, offering a versatile tool for educational, therapeutic, and entertainment purposes. Users can customize various elements such as the hero, setting, theme, genre, and illustration style to create unique stories tailored to a child's individual preferences or specific needs, like addressing emotions or promoting desired behaviors. The platform also provides advanced text modification options, including highlighting, adjustable reading pace, font size and type changes, and color selection, making it suitable for children with dyslexia or for use in speech therapy and auditory processing training.

The Smart Data Analytics (SDA) research group, supported by Prof. Dr. Jens Lehmann, is a virtual research group with researchers from TU Dresden and the Institute for Applied Computer Science, along with external PhD students. Their work intersects machine learning and knowledge graphs, covering areas such as question answering, dialogue systems, and representation learning for knowledge graphs. The group develops horizontally scalable analytics algorithms for large-scale knowledge graphs, focusing on machine learning over knowledge graphs by computing embeddings and learning description logic concepts. They also research semantic question answering using natural language processing and software engineering for data science, aiming to align data and software engineering methods.

QuikAuthor is an AI-powered microlearning platform designed to accelerate the creation of impactful e-learning content. It leverages artificial intelligence to transform various content formats, including videos, PDFs, and text prompts, into interactive courses. The platform focuses on driving engagement and knowledge retention through gamification elements. Users can efficiently generate complete courses, incorporating quizzes, assessments, and engaging activities. QuikAuthor supports uploading existing educational materials or simply describing the desired teaching content, making course development accessible and fast. It also offers the capability to export courses as SCORM files, ensuring compatibility with various learning management systems.

Any Diffuse is an AI image generation tool hosted on Hugging Face Spaces, enabling users to create high-quality images by simply entering text prompts. The platform provides various customization settings, allowing users to select different models, styles, and resolutions to achieve their desired visual output. While the tool aims to offer a flexible and accessible image generation experience, it is currently experiencing runtime errors, preventing full functionality. Despite the technical issues, its design suggests a focus on user-controlled image creation through stable diffusion technology.

Generative-AI-with-LLMs is a comprehensive learning resource designed to equip developers with a deep understanding of generative AI and Large Language Models (LLMs). The course covers the entire generative AI lifecycle, from data gathering and model selection to performance evaluation and deployment in real-world applications. Learners will explore the transformer architecture, LLM training, fine-tuning techniques, and empirical scaling laws for optimizing models. It also delves into state-of-the-art training, tuning, inference, tools, and deployment methods, while discussing the challenges and opportunities generative AI presents for businesses through industry insights. The curriculum includes practical labs and quizzes to reinforce learning.