5G INSTRUCT forge: A data engineering pipeline for making LLMs learn 5G

Large Language Models (LLMs) have transformed various fields with their remarkable ability to comprehend and generate human-like text. Despite these advancements, their effectiveness in specialized domains such as finance, law, medicine, and telecommunications remains limited. To adapt these models to new domains, it is essential to train them on relevant datasets. Fine-tuning is a well-known method for training LLMs on new tasks using specialized datasets. However, generating these specialized datasets presents a critical challenge, as structuring the data appropriately for effective learning is complex. To address this challenge, this paper presents 5G Instruct Forge, an advanced data engineering pipeline designed to create domain-specific datasets for 5G networking, particularly from the 3rd Generation Partnership Project (3GPP) specifications. By processing unstructured documents, i.e., 3GPP Technical Specifications (TSs), into structured formats, our pipeline enables LLMs to be fine-tuned for understanding and generating 5Grelated content. As a proof of concept, we generated the OpenAirInterface

(OAI) Instruct dataset using our pipeline, utilizing a subset of the 3GPP TSs used to develop OAI. Evaluation results demonstrate that training generic open-source LLMs on this dataset resulted in new 5G-aware LLMs outperforming OpenAI’s GPT-4 on 5G-specific tasks.