aestudio/Olmo-3-7B-Instruct-DPO-GRPO-math-lora-gen270

Held-out test result: 0.699 vs RLVR 0.702, parity (-0.003 [-0.012, +0.006])

Retention measured on this checkpoint: IFEval flat (+0.006), ARC +0.063

LoRA adapter: grpo_r2_ext_s0 checkpoint-480 = combined generation 270 (sha256 dcb0ef48…)

Model card for the generation-270 adapter

initial commit