This paper introduces a novel benchmark for personalized text generation with fine-grained linguistic control. The benchmark evaluates the ability of generative models to generate text that reflects specific linguistic attributes, such as lexical, syntactic, and rhetorical features. The authors use multiple datasets, including blogs, movie reviews, and product reviews, to create a diverse set of examples for training and evaluation. They extract various linguistic attributes from the text, such as token count, sentence count, readability, part-of-speech tags, dependency relations, and rhetorical structure. These attributes are then discretized to enable controlled modification of specific linguistic features. The authors evaluate the performance of various large language models, including the Pythia series and GPT-3.5, on their benchmark. They find that larger models generally perform better, and that the 1B Pythia model achieves the best results. The study also investigates how models respond to changes in linguistic attributes, revealing that performance varies depending on the attribute and its value. The authors also explore the impact of training data size on model performance, finding that performance decreases as the number of training examples per author decreases. The paper highlights the importance of fine-grained control in text generation, as it allows for more accurate and personalized text generation. The authors propose a benchmark that enables researchers to evaluate and improve models for personalized text generation. They also provide insights into the factors that influence model performance, such as the number of attributes, the variation in attribute values, and the model's ability to learn contextual representations. The authors make their code, data, and pretrained models publicly available to encourage further research in this area.This paper introduces a novel benchmark for personalized text generation with fine-grained linguistic control. The benchmark evaluates the ability of generative models to generate text that reflects specific linguistic attributes, such as lexical, syntactic, and rhetorical features. The authors use multiple datasets, including blogs, movie reviews, and product reviews, to create a diverse set of examples for training and evaluation. They extract various linguistic attributes from the text, such as token count, sentence count, readability, part-of-speech tags, dependency relations, and rhetorical structure. These attributes are then discretized to enable controlled modification of specific linguistic features. The authors evaluate the performance of various large language models, including the Pythia series and GPT-3.5, on their benchmark. They find that larger models generally perform better, and that the 1B Pythia model achieves the best results. The study also investigates how models respond to changes in linguistic attributes, revealing that performance varies depending on the attribute and its value. The authors also explore the impact of training data size on model performance, finding that performance decreases as the number of training examples per author decreases. The paper highlights the importance of fine-grained control in text generation, as it allows for more accurate and personalized text generation. The authors propose a benchmark that enables researchers to evaluate and improve models for personalized text generation. They also provide insights into the factors that influence model performance, such as the number of attributes, the variation in attribute values, and the model's ability to learn contextual representations. The authors make their code, data, and pretrained models publicly available to encourage further research in this area.