Influcoder: Distilling Decoders' Gradient Influence Rankings into an Encoder for Data Attribution
This work addresses the scalability bottleneck of influence-based data attribution for practitioners curating large training datasets for LLMs, offering a practical solution for identifying problematic samples like sources of toxic behavior.
Influcoder proposes a method to accelerate influence-based data attribution for large language models by distilling gradient influence rankings from a decoder into an encoder, enabling faster and more storage-efficient processing. The approach achieves comparable attribution accuracy to traditional influence functions while significantly reducing computational overhead.
With the growth of LLMs' (Large Language Models) capabilities, there has been an increasing push to curate high quality datasets by filtering samples in the training data. In general, Data Attribution (DA) methods aim to estimate how individual samples in a training dataset can precondition a model to generate certain outputs. As an example, one might be interested in which samples in the data could be the source of toxic behavior after training the LLM. Many methods quantify this conditioning through the paradigm of influence functions. While methods of this family are effective in its function, they lack the necessary processing speed and storage compactness to be practically implemented on large datasets. We propose a method, Influcoder, as a quick and cost-effective approach to influence-based Data Attribution at scale.