From selected queries in the 90K pool, we generated cited reports. Quality filtering left 39.5K examples for supervised fine-tuning (SFT).
We used separate queries to build ~6K report pairs where two judge models agreed, then trained with direct preference optimization (DPO).
