Customize small language models on AWS with automotive terminology

This article discusses the process of customizing small language models (SLMs) for specialized domains, specifically focusing on automotive diagnostics using AWS services like Amazon SageMaker and Amazon Bedrock. The key highlights of the approach include:

• Using the Automotive_NER dataset to fine-tune a Meta Llama 3.1 8B Instruct model
• Employing advanced data selection techniques using TF-IDF vectorization to create a balanced training dataset
• Applying Low Rank Adaptation (LoRA) for efficient model fine-tuning
• Demonstrating two deployment methods: SageMaker real-time inference and Amazon Bedrock Custom Model Import
• Evaluating the fine-tuned model using BLEU scores and Normalized Levenshtein distance

The evaluation showed that the fine-tuned model outperformed base models in understanding automotive terminology and generating more accurate diagnostic responses. The approach can be extended to other specialized domains, showcasing the potential of customizing small language models for specific use cases.