The J-lens offset is the model's token frequency: z-scoring helps

Researchers found that the J-lens offset in language models corresponds to token frequency, and applying z-scoring improves its interpretability and usefulness.