Bertinoro International Spring School 2024

Large Language Models and How to Instruction Tune Them (in a Sustainable Way)

Authors: Danilo Croce

Many thanks to: Claudiu Daniel Hromei for supporting the development of (most of the) code

This repository hosts materials from the Bertinoro International Spring School - BISS-2024 tutorial.

The objective of this tutorial is:

• Introduce the Basics of Distributional Semantics, and the interplay with neural learning.
• Introduce Transformer-based architectures, including encoding-decoding, encoder-only, and decoder-only structures.
• Demonstrate fine-tuning of Large Language Models (LLMs) on diverse datasets in a multi-task framework.
• Utilize Low-Rank Adaptation (LoRA) for sustainable and efficient tuning on "modest" hardware (e.g., single 16GB RAM GPU).

The repository includes code for fine-tuning a Large Language Model (based on BERT and LLaMA) to solve NLP tasks, such as the ones proposed in EVALITA 2023.

Code

Lab 1: Training BERT-based models in few lines of code

This is a Pytorch (+ Huggingface transformers) implementation of a "simple" text classifier defined using BERT-based models. In this lab we will see how it is simple to use BERT for a sentence classification task, obtaining state-of-the-art results in few lines of python code.

The python book is available at this LINK.

Lab 2: Fine-tune a LLaMA-based model for all tasks from EVALITA 2023

At the end, this tutorial shows how to encode data from different tasks into specific prompts and fine-tune the LLM using Q-LoRA. The code can be also used in Google Colab using an Nvidia-T4 GPU with 15GB memory.

The code is heavily based on the one used in ExtremITA system participating to EVALITA 2023:

• ExtremITA Paper
• ExtremITA Github Code

The overall process is divided in four steps:

• Step 1 - Encoding the data: it shows how to encode data from an EVALITA task to generate prompts for the LLM
• Step 2 - Fine-tuning the LLaMA model: it shows how to fine-tune the LLMS given the prompts
• Step 3 - Inference: generating answers: it shows how to use the fined-tuned model
• Step 4 - Deconding the data: it shows how to conver the data to be evaluated in the EVALTA challenge

Slides

The repository also features tutorial slides (LINK).

Project Assignment

The process assignment is presented at the following LINK.

Contacts

For queries or suggestions, raise an Issue in this repository or email croce@info.uniroma2.it