Amazon SageMaker Canvas and Amazon SageMaker JumpStart have brought a new level of accessibility to fine-tuning large language models (LLMs), allowing businesses to tailor customer experiences precisely to their unique brand voice. No coding is needed for this process, as SageMaker Canvas provides a user-friendly, point-and-click interface. This not only allows faster operation but also uses fewer technical resources.
Fine-tuning LLMs on company-specific data enhances uniform messaging across various customer touchpoints. SageMaker Canvas empowers businesses to personalize customer experiences without the need for extensive technical know-how. Moreover, all business data used in the process is kept within a secure AWS environment, and is neither used to improve base models nor shared with third-party model providers.
Any first-time user of this process would require an AWS account and an AWS Identity and Access Management (IAM) role with SageMaker and Amazon Simple Storage Service (Amazon S3) access. Following this, users need to create a SageMaker domain—a collaborative ML environment with shared file systems, users, and configurations.
A dataset, specifically a prompt/completion pair file in CSV format, is also necessary for the supervisory fine-tuning undertaken by SageMaker Canvas. This allows the system to learn how to provide specific responses with the appropriate format and adaptation.
The process of creating a new model involves selecting ‘My models’ in SageMaker Canvas’ navigation pane and then choosing ‘New model’. Here, the user can enter a model name and select the problem type as ‘Fine-tune foundation model’.
Once the new model has been created, the dataset can be imported into SageMaker Canvas, which scans it for any formatting issues. The user can then select the foundation model (FM) of choice and proceed with fine-tuning in line with the dataset.
The entire process can take between two to five hours, with SageMaker splitting the dataset into an 80/20 split for training and validation. Once this process is complete, the user can review the new model’s stats and generate an evaluation report for better understanding and adjustment.
Fine-tuning isn’t the ultimate solution for every need. Other approaches such as prompting, retrieval augmented generation (RAG) architecture, continued pre-training, postprocessing, and fact-checking can all be used in concert with fine-tuning to create the ideal AI solution.
By using fine-tuned LLMs, organizations can effectively create an AI that speaks in their brand’s voice, thus enhancing customization at all customer touchpoints. The system is robust and versatile enough to be used in various real-world applications, provided the organization’s datasets are in CSV format. The opportunities are limitless if organizations also consider the benefits and trade-offs of different approaches.
