2023 saw the incredibly rapid development of generative artificial intelligence (AI) technologies, based on large language models (LLM). The development of large language models, especially multimodal (text, image, sound, programming code) has accelerated the possibilities of application in various business segments such as marketing - managing sales channels and general communication with customers and customers, writing minutes and summaries of large business studies; in medicine for faster writing of medical reports and diagnoses, in education to help with knowledge verification; in law for faster analysis of legal files and writing legal documents, in scientific research for anticipation of new antibiotics and others. Studies show that generative AI has already brought changes to the labour market and that changes will continue at a rapid pace.
However, the economic benefits of applying this technology come with a number of Potential safety issues such as the generation of inaccurate content, referred to as hallucinations, bias and toxicity, the reliability of the generated text, copyright and privacy. These risks can significantly damage the company's reputation and affect society.
EDIH Adria project has prepared a series one-day workshops where you can find the best way to increase your own productivity and grow your business through generative AI.
Prompt engineering so covers topics like: Introduction to generative AI and impact on business and labour market. An overview of the tasks that generative AI can solve and which challenges it is not up to with an overview of risks and limitations. It will answer the question of what is the difference of using pre-trained and fine-tuned models, and will give an overview of models for generating text, images, speech, music and video with examples of applications in marketing, medicine, business, etc. Further sitematizing will be prompt engineering approaches and recommendations with Zero-shot Prompting, Few-shot Prompting, Chain-of-Thought Prompting, Tree of Thoughts prompting, ReAct prompting, Active-Prompting, Multimodal CoT prompting, etc.
Introduction to the development of LLM-based applications (large language model) is an education based on the application of the Deep Grid Transformer architecture, i.e. its decoder part. In the introductory part, the principles of learning large language models for text generation, their fine-tuning and the principles of evaluation on benchmark data will be presented. Also, techniques for the application of large language models in applications such as fine tuning, the problem of castastrophic forgetting and learning within the context will be presented. Parameter Efficient Fine-Tuning-PEFT, adapter and Low-Rank Adaptation of Large Language Models-LoRA techniques will be presented. Particular attention will be paid to generating factually accurate texts with Retrieval Augmented Generation (RAG) techniques using vector bases, knowledge graphs or external knowledge. LLM agents and the most important software packages for building applications such as LangChain and LlamaIndex will be explained. After this course, users will distinguish between pre-trained and fine-tuned models, and be able to start developing applications that use generative AI responsibly and safely.
If users want to try their hand at building your own LLM on BERT architecture, Advanced education covers basic architectures and deep learning principles with activation and loss functions. The differences between the recurrent network and the transformer network, as well as its encoder, decoder and encoder-decoder variants will be explained. At the core of the architecture are the mechanisms of attention Self attention and Multihead attention. The principles of autoregressive and masked learning will be explained, as well as the difference between static and dynamic (contextualized) vectors for the representation of languages - embeddings. The BERT model comes with variations: RoBERT, AlBERT, SpanBERT. For model learning to be successful, it is important to understand the principles of transfer learning by fine-tuning the model to the user's task from the domain of sequence prediction, classification or sentence generation.
English 
Croatian 
Italian