How language models judge the odds of an event: data, code and pre-registrations

Data, code and pre-registration documents for the manuscript How language models judge the odds of an event (Wenbin Wu, Cambridge Centre for Alternative Finance, University of Cambridge). The archive holds the scenario lists and both wordings; the window-frequency table with its rules and source series; the corpus mention counts and the stated-probability census over the Dolma-3 mix; the per-checkpoint readouts for OLMo-3 and the cross-family panel; the steering directions and ablation results; the continued-pre-training corpora with their audits; the twenty pre-registration documents with their hashes; and the scripts that produce the numbers, tables and figures of the paper. Model weights and the Dolma-3 corpus are the public releases of their developers and are not included. Code is released under the MIT licence; data, corpora and text under CC BY 4.0.