MiDRASH Geniza 01 HTR model

This first MiDRASH Geniza model was trained in the frameworks of the ERC Synergy project MiDRASH (Grant 101071829) at the EPHE, PSL in Paris and the HTR4PGP project DataXfund at the University of Princeton and with support from the ERC MAJLIS (Grant No. 101002243) at the LMU Munich. It was used in the production of the Hebrew script items in our first complete transcription of the Cairo Genizah (https://zenodo.org/records/17734473). We invite people interested in transcriptions of Hebrew manuscripts to submit their requests via this form: https://forms.gle/pDDFmT5xDZmKKFrv9 or to publish any transcription created with it in an open repository so that we can further improve the model. The model is the result of finetuning on documentary and literary texts (Tannaitic literature, Talmud and some Judeo-Arabic) from the Cairo Genizah in Judeo-Arabic, Aramaic and Hebrew on top of a basemodel trained on a huge variety of Hebrew script texts and prints. Abbreviations are not resolved. Transcriptions follow the MiDRASH guidelines (https://zenodo.org/records/17931140). Parentheses indicate inline or single letter supralinear additions. Doublesquare brackets indicate deletions. Square brackets indicate lacunae and linefillers. The model was trained with NFKD unicode normalization. More details will be furnished in a publication in preparation.