Compared to off the shelf GenAI models such as the GPT series from OpenAI – the obvious challenge is that the fine-tuned model needs to be validated and checked for errors that are derived from data quality issues. Here is a list of some of the common types of risks:
Resource Requirements: Depending on your ambitions, it can take a long time. The higher quality required, the larger the volume of data to be produced or checked it types more resources and time. It’s very important to think about how to fit this into your existing processes and to think about how your business continuously improves the data.
Overfitting to the Fine-tuning Dataset: The model may perform exceptionally well on the fine-tuning data but poorly on unseen, real-world data that slightly differs. This is because it memorizes the training examples rather than generalizing the underlying principles.
Forgetfulness: The model ”forgets” knowledge acquired during pre-training when fine-tuned on a narrow dataset or for a specific task.
Bias Amplification: Biases present in the fine-tuning dataset can be amplified, leading to discriminatory or unfair outputs.
Privacy Leakage: If the fine-tuning dataset contains sensitive information, the model may inadvertently memorize and leak it through its outputs.
Difficulty of Control: It can be challenging to precisely control the behavior of a fine-tuned model and ensure it aligns with specific requirements or constraints. Thus, it’s better suited for narrow use cases rather than general use.