wassname/

ultrachat_200k_ultraclean

A sycophancy-filtered subset of wassname/ultrachat_200k_filtered (itself a refusal-filtered derivative of HuggingFaceH4/ultrachat_200k). Built as a clean SFT base for studying sycophancy-reduction methods (DPO, antipodal LoRA, etc.). We want the SFT base to NOT already have reflexive validation baked in, so the downstream preference-optimization signal is not competing with an SFT prior that already says "Great question!" by default.