A preference dataset targeting hedging and evasion in LLM answers to empirical questions. chosen responses lead with a magnitude and name their sources; rejected responses deliver the same claims with the information stripped out. 144 pairs (130 train / 14 test). Generated with Grok (xAI) grok-4.5 with web search. This is a pilot — large enough to validate the pipeline and run a small DPO experiment, not large enough to move a model on its own.

The core… See the full description on the dataset page: https://huggingface.co/datasets/nbeerbower/weasel-dpo.