<ul><li>Training data attribution (TDA) methods aim to identify influential training examples for a model's predictions on specific test data.</li><li>Gradient-based TDA methods are limited in scalability, and recent random projection-based methods often sacrifice attribution accuracy.</li><li>Daunce is introduced as a data attribution approach through uncertainty estimation, fine-tuning perturbed models and computing covariance of losses for attribution scores.</li><li>Daunce is scalable to large language models, achieves more accurate attribution, and successfully applied to OpenAI's GPT models for data attribution.</li></ul>

Daunce: Data Attribution through Uncertainty Estimation

Discover more