Preference-trained OLMo-3 checkpoints from the sycophancy-confidence project.
Each top-level folder is one training run (HF transformers-loadable weights). DeepSpeed global_step* / step_* resume state is omitted.
transformers
global_step*
step_*
-
