The TAO of data: How Databricks is optimizing AI LLM fine-tuning without data labels
1 min read
Summary
AI models can only be as good as the data used to train them, making the collection of relevant, labelled data a problematic bottleneck when implementing AI applications.
Databricks’ new Test-time Adaptive Optimization (TAO) approach aims to remove this bottleneck by enabling enterprises to tune large language models using only their existing input data, with no labelled data required, and achieve better results than with traditional fine-tuning methods.
Test-time compute is not a new idea, but TAO differs in that, while using additional compute during training, the final tuned model has the same inference cost as the original model, offering an advantage for production deployments where inference costs scale with usage.
Databricks claims TAO outperforms traditional fine-tuning in multiple enterprise-relevant benchmarks, offering a cost-saving and time-saving opportunity for technical decision-makers to implement AI capabilities faster and more efficiently.