Онлайн събитие

Rearchitecting LLMs: Building Specialized LLMs for Real Projects

Лектор Pere Martra

Организатор

Job Board за
IT общността

Дата 13.05.26

Език English

Регистрирай се

Спонсори на събитието:

За събитието

*The event will be in English.

This event is part of the AI/ML/Big Data User Group.

AGENDA:

Current AI applications need models that are fast, small, and domain-specific. This session opens the black box and shows how to structurally modify LLMs from the inside, adapting them to specific data and goals.

>>> Real-world cases: why specialized LLMs matter (BMW, BlackRock, JPMorgan)
>>> Depth Pruning: removing layers intelligently using PyTorch hooks and cosine distance
>>> Width Pruning: Explaining (GLU architecture) and adapting Mistral Ministral 3 paper to our needs.
>>> Knowledge Distillation: recovering model performance after structural modification
>>> OptiPfair: open-source library encapsulating the full pipeline
>>> FairPruning: fairness-aware pruning for domain-specific models

>>> Q&A част

The event is free, but pre-registration is required.

USER GROUP: AI/ML/Big Data

The event is part of the AI/ML/Big Data User Group. Every month we organize events on a topic related to AI/ML/Big Data. In order not to miss interesting information on the subject – subscribe.

ORGANIZED BY: DEV.BG is a specialized Job Board for IT jobs.

AI/ML/Big Data jobs: If you are looking for a job for AI/ML/Big Data, view the job listings here.

WHEN: May 13th, 2026, 07:30 pm.

Pere Martra AI Research Engineer & Author

За лектора

Pere Martra is an AI Research Engineer specializing in LLM optimization, structural pruning, and fairness-aware compression. He is the author of Large Language Models Projects (Apress, 2024) and Rearchitecting LLMs (Manning, MEAP 2026), a book covering state-of-the-art techniques to modify and optimize LLM architectures.

His research focuses on structural pruning, knowledge distillation, and fairness-aware optimization. He published the arXiv preprint „Fragile Knowledge, Robust Instruction-Following: The Width Pruning Dichotomy in Llama-3.2“ and created OptiPfair, an open-source library implementing the full pruning and distillation pipeline.

Pere maintains GitHub repositories with nearly 2,000 stars, and his work has been adopted as reference material at PES University (B.Tech CSE AI & ML) and the University of Texas at Dallas.