Seminarium: Efficient Small-scale Large Language Models

Zapraszamy na seminarium naukowe Pracowni Inteligentnych Systemów Wspomagania Decyzji (ISWD) Politechniki Wrocławskiej, Zakładu Wspomagania Decyzji w Warunkach Ryzyka IBS PAN oraz Centrum Technik Informatycznych WIT, które odbędzie się 25 października 2023 (środa) o godzinie 13:15 w formie online.

Podczas seminarium dr inż. Maciej Hojda z Pracowni ISWD wygłosi referat pt. Efficient Small-scale Large Language Models.

Abstract
Language modeling is a way to equip machines with so-called language intelligence. In general, the aim is to model the generative likelihood of token sequences (typically word sequences) to predict the probabilities of future tokens. Recently, a new type of paradigm, called large language models (LLM), has emerged on the foremost front of language modeling. An LLM is a transformer-type neural network that takes in some text in natural language and outputs another text. LLMs are a type of language model for which the rise in complexity counted by the number of nodes in the neural network (in tens and thousands of billions) led to the development of abilities previously unobserved, such as in-context learning, instruction following, and step-by-step reasoning.
Recently, new models have been made progressively more accessible to hobbyists and the general public alike – high-quality LLMs such as Llama 2 can be efficiently run on a typical consumer machine. This presentation is a primer to LLMs in general and to Llama 2 models in particular, where the latter's efficiency is evaluated for specific generation parameters.

Seminarium odbędzie się 25 października 2023 (środa) o godzinie 13:15 za pośrednictwem platformy Zoom  Link do spotkania.