What it does This hub documents an academic project building 4B-parameter language models trained from scratch exclusively on texts published before hard cutoff dates: 1913, 1929, 1933, 1939, and 1946. The upcoming Ranke-4B family uses Qwen3 architecture, 80B training tokens per model, and a curated pool of 600B time-stamped tokens. The goal is giving historians and social scientists an interactive aggregate witness to each era’s published discourse—not a modern AI pretending to wear period costume.

The interesting bit The researchers call their approach “uncontaminated bootstrapping”: they build chatbots while deliberately not correcting the models’ normative judgments. Ask the 1913 model about Hitler and it invents a harmless German academic, because that name hadn’t accumulated meaning yet. The project treats reproduced racism, misogyny, or imperialism as features for study, not bugs to sanitize—understanding how such views were articulated requires letting the model voice them.

Key highlights

• Five knowledge-cutoff dates spanning the interwar period and WWII
• 4B parameters, Qwen3 architecture, trained from scratch on historical corpora
• Explicitly not aligned to modern values; designed for scholarly interrogation of historical discourse
• Companion repos for pretraining, data curation, and posttraining listed but not yet live
• Models and datasets planned for Hugging Face release; working paper forthcoming

Caveats

• The actual model weights, code, and data repositories are still “coming soon” as of the README date
• The authors note these are tools for exploring published text, not perfect mirrors of public opinion or popular belief
• Access for the broader public may require a “protective layer” against extreme toxic outputs, though details are unspecified

Verdict Historians, digital humanists, and anyone studying ideology or discourse should watch this closely. If you need a production chatbot or want sanitized historical fiction, look elsewhere—the whole point is the unsanitized specificity.