Pre-1900 family archives of hand-written letters donation - how to organize?
Intended downstream use is corpus generation — manufacturing period-consistent pretraining data ("VictorianWikiText") for larger 19th-century models, since the supply of genuine clean period text is largely exhausted.
Can researchers organize a relyable supply? What motivation can support our histrorical LLM research? Same as with Genealogy - private efforts are less effective, than collective.
I don’t think it’s about organizing a reliable supply there just isn’t enough period. Talkie 13B worked from 1930 prior— for the Victorian era if we clean up all the OCR and take every written letter and every written book and pretrain off of it we are still going to have a 10B model or less…
This is why this experiment is in part to evaluate ways to synthetically augment this data in a form that is going to be better for instruction tuning.