On Training Data Influence of GPT Models
This addresses the need for better interpretability of training data effects in large language models, though it appears incremental as an extension of influence estimation methods to GPT architectures.
The paper tackles the problem of understanding how training data influences GPT model performance by introducing GPTfluence, a featurized simulation approach that traces individual training examples' impact on performance metrics across models with 14 million to 2.8 billion parameters. The method demonstrates robust generalization to unseen data in fine-tuning and instruction-tuning scenarios.
Amidst the rapid advancements in generative language models, the investigation of how training data shapes the performance of GPT models is still emerging. This paper presents GPTfluence, a novel approach that leverages a featurized simulation to assess the impact of training examples on the training dynamics of GPT models. Our approach not only traces the influence of individual training instances on performance trajectories, such as loss and other key metrics, on targeted test points but also enables a comprehensive comparison with existing methods across various training scenarios in GPT models, ranging from 14 million to 2.8 billion parameters, across a range of downstream tasks. Contrary to earlier methods that struggle with generalization to new data, GPTfluence introduces a parameterized simulation of training dynamics, demonstrating robust generalization capabilities to unseen training data. This adaptability is evident across both fine-tuning and instruction-tuning scenarios, spanning tasks in natural language understanding and generation. We make our code and data publicly available at https://github.com/ernie-research/gptfluence.