r18-a2-q2-u250-r2f2-s300
beatrizh/r17 + quality RFT round 1: SFT on its own answers that beat king-2025 (both judge orders, gemini-2.5-flash-lite) w-sft 2.0, lr 2e-06 last:8, 2011 known protection (LLM-known + 48k clozes), <=2011 text KD 0.0 / CE 0.0, no forget, step 300