LLM Fine-Tuning
Fine-tuning a small LLM on a domain-specific knowledge base.
Requirements
- M.Sc. in Machine Learning, Data Science, Computer Science, Mathematics, Telecommunications, or a related field
- Strong knowledge of Python
- Familiarity with Large Language Models (LLMs)
- Understanding of model training concepts, including fine-tuning, overfitting, layer training, and catastrophic forgetting
Description
In this internship, the candidate will investigate the fine-tuning of a small Large Language Model (i.e., fewer than 3B parameters) on a domain-specific knowledge base that is largely disjoint from the model’s original training data.
The main objective is to enable the LLM to acquire new domain knowledge while minimizing catastrophic forgetting—that is, without significantly degrading previously learned capabilities. The internship will focus on analyzing the trade-off between knowledge acquisition and retention.
A custom dataset will be selected for domain-specific training, while standard benchmark datasets will be used to evaluate the preservation of the model’s original knowledge throughout the fine-tuning process.