 And we straight move on to Susanna Tavares-Pedro from the University of Lisbon, which again should be an online presentation if I'm not wrong. Sharing my screen now, so are you able to see the presentation? Yes. Okay, thank you very much. So this is the quick presentation assessment of the last nine months of the Traprinck project, or TPQ for short, which is an 18 month exploratory project with a grant from the Portuguese Foundation for Science and Technology, ending in July 2023. The team has 11 transcribers, both paleographers and historians and five consultants. TPQ's goal is to create a generic HDR plus model with at least one million words in order to obtain a maximum of 5% CER invalidation test. We will also be preparing the framework for the next phase, which will have two main components, the full transcription of inquisition court proceedings using the HDR model and their subsequent XMLTI edition. The model is built from digitized images of the 18,000 judicial proceedings from the Lisbon Tribunal of the Holy Inquisition, which are available online on the Torre do Tombo National Archives catalog website. This is a large scale model for several reasons. First, the quantity of pages to transcribe to reach the goal of one million words, we estimate 5,000 pages to transcribe. Then the wide chronological time span of the collection from 1536 to 1821, 285 years spanning four centuries, which means that we have to choose our images from something between two to four million pages, just from the Lisbon court file series alone. The variety of scripts and individual hands are enormous, ranging from very fast Gothic cursive from the 16th century to calligraphic and pancellarist humanistic scripts to early 19th century personalized modern hands. This means that we need to have ground truth from as many hands as possible in an estimated universe of three to 5,000 different scribes. All these factors are challenges to which can be added the transcribers initial difficulty in making the sort of transcription required to work with transcripts very different from their usual practice. So we are now nine months into the project and have just completed the second training of the model, both with pile layer and HDR plus engines, which were fed some 20,000 and 35 pages, about 400,000 words with very promising results already. The model has a 7.3% CER invalidation set and the HDR plus model has 6.7 CER. From our experience so far, these are the main issues to address right from the start in order to minimize the challenges of such large scale model training. First, you have to know the texts you will be working with and put together a comprehensive set of transcription guidelines, making sure everyone on the team is acquainted with it and update it whenever an unforeseen situation requires any rule to be added. Practice transcribing with the selected criteria and always have someone else to review your work. Share all your doubts and difficulties in reading the manuscripts in a forum also open to the rest of the team in order to solve problems quickly and as a group, we use Slack which has been working very well with us. Try to have balanced training and validation ground proof sets so they represent accurately the variety of hands in all the different decades. And take a vague advantage of the tools transcribers offers for model performance analysis to add more samples from those hands where the model perform less well to train and then train again. Thank you all very much. Good luck and good work. Bye bye. Thank you. Perfect. Thank you very much for this presentation as well.