Native-space, per-sentence chain-of-thought activations for 6 reasoning models on medical case
diagnosis (MedCaseReasoning: medqa + nejm, and medmcqa). Each model's responses were regenerated
with its own native chat template (greedy decoding), then teacher-forced through the model to
extract a hidden-state vector for each reasoning sentence.
responses/ full regenerated responses… See the full description on the dataset page:
https://huggingface.co/datasets/matthewshu/medcasereasoning-cot-activations.
