Scaling Laws for Linear Complexity Language Models
This work addresses the scalability of linear complexity models for language modeling, which is incremental as it builds on existing architectures to confirm their performance.
The study investigated the scaling capabilities of linear complexity language models, finding that they exhibit similar scaling to transformer-based models and show superior linguistic proficiency and knowledge retention, based on training models from 70M to 7B parameters on a 300B-token corpus and evaluating 1,376 checkpoints.
The interest in linear complexity models for large language models is on the rise, although their scaling capacity remains uncertain. In this study, we present the scaling laws for linear complexity language models to establish a foundation for their scalability. Specifically, we examine the scaling behaviors of three efficient linear architectures. These include TNL, a linear attention model with data-independent decay; HGRN2, a linear RNN with data-dependent decay; and cosFormer2, a linear attention model without decay. We also include LLaMA as a baseline architecture for softmax attention for comparison. These models were trained with six variants, ranging from 70M to 7B parameters on a 300B-token corpus, and evaluated with a total of 1,376 intermediate checkpoints on various downstream tasks. These tasks include validation loss, commonsense reasoning, and information retrieval and generation. The study reveals that existing linear complexity language models exhibit similar scaling capabilities as conventional transformer-based models while also demonstrating superior linguistic proficiency and knowledge retention.