zimplex/olmo7-think-sft-euc

Thinking-SFT of allenai/Olmo-3-1025-7B on allenai/Dolci-Think-SFT-7B (standard euclidean head), trained with the <think> ChatML template. This is the capacity-matched euclidean control for the hyperbolic-head experiment.

Usage

Load normally: AutoModelForCausalLM.from_pretrained("zimplex/olmo7-think-sft-euc"). Prompt with the <think> template; greedy decoding. Take the answer after </think> (last \boxed{} / ####), graded by math_verify.