Marco-o1-CoT相比基线模型Qwen2-7B-Instruct提升了1.37% Marco-o1-MCTS (step)进一步提升至90.40%,比基线模型提高了显著的6.17% 在中文MGSM数据集上: Marco-o1-MCTS (mini-step of 32 tokens)达到82.40%的准确率,比基线模型提高了令人瞩目的5.60% 这些结果清楚地表明,通过巧妙结合CoT微调、MCTS搜索和创新的推理...
An Open Large Reasoning Model for Real-World Solutions - Update CoT_demo.json · AIDC-AI/Marco-o1@fa821e2