SmolLM2-135M-Reasoning SLERP ARC-E Champion (v2)
NEW ALL-TIME ARC-E RECORD: 59.51%
SLERP merge (t=0.2) of exp024 (SFT 1K) + exp136 (DPO IPO loss).
ARC-E=59.51%, PIQA=67.08%, HellaSwag=43.14%, ARC-C=28.84%.
Key insight: IPO model is a more complementary merge partner than sigmoid.