Measuring Data Influence in LLM Training

Summary & Methodology Analysis

The researchers propose a novel way to measure data influence that does not rely on specific downstream tasks, which have historically made such measurements difficult. Their approach defines the influence of a mini-batch as the reduction in the squared L2 distance between the current model parameters and the final pretrained parameters. By decomposing this influence into individual example contributions and utilizing saved intermediate checkpoints, they enable an analysis that avoids the prohibitive cost of full retraining. This allows engineers to audit the training process retrospectively by evaluating how specific chunks of data shaped the model's ultimate behavior.

The study evaluates this method using the Pythia model family and The Pile dataset. The results indicate that the importance of data shifts throughout the pretraining phase. For the Pythia-1.4B-Deduped model, the researchers observed that the mean contribution of examples peaks at approximately 40k training steps. Furthermore, they identified a category of data they label as opponent examples, which push the model away from the final learned configuration. The prevalence of these opponent examples reaches 8 to 10 percent by the end of the training cycle.

There are two primary limitations to consider when applying this approach. First, the influence metric is relative to a specific pretraining run, meaning that choosing different reference checkpoints will change the resulting rankings. Second, while the mathematical decomposition of influence is derived from stochastic gradient descent (SGD), the Pythia models analyzed actually employ adaptive optimizers. This discrepancy between the theoretical derivation and the practical implementation should be taken into account when interpreting results in environments using modern optimization strategies.