A prompt-trained DeepSeek R1 70B can perform better than GPT-o1 using AdalFlowcolab.research.google.com4 pointsmeame2010a year ago