Vast amounts of medieval notarial documents have been preserved and digitized in Catalan archives. These collections provide invaluable insights into life in the Middle Ages. To harness their full potential, documents need to be accurately transcribed. This paper introduces a new dataset for the task of handwritten text recognition on medieval manuscripts, focusing on notarial charters written on parchment from the 13th to 15th centuries. Our dataset is comprised of 100 digitized manuscripts (3,369 lines), carefully selected to represent the large variation that is present in the sources, encompassing at least 80 distinct hands, various document types (from sales and inventories to last wills and marriage contracts), and spanning three centuries. Written primarily in Medieval Latin with fragments in Medieval Catalan, these manuscripts exhibit varying stages of preservation and degrees of deterioration, resulting in a very diverse dataset. Recent years have seen impressive advancements in automatic handwriting recognition of historical manuscripts, with significant effort put into creating general models that perform well across several centuries and scripts. The experiments on our dataset show that these models are still far from providing a generalizable solution, but nevertheless provide a good foundation to fine-tune robust models when used in combination with a carefully curated dataset. We report a series of experiments using an extensive and nuanced evaluation approach that demonstrates the importance of data selection for achieving an accurate transcription of historical manuscripts.