I am very glad to see such an excellent work!
Unfortunately, I seem unable to reproduce the results in the paper. My environment is CUDA 12.2, vllm 0.6.6, and I used the validation script to select the checkpoint (step 90). Regrettably, I still failed to reproduce the paper's results. Even when I tested your released Qwen-math-2.5-CFT in the exact same environment, there was a significant difference from my trained model.
In particular, the numerical values on AMC23 differ by nearly 10 points. I wonder how I can successfully reproduce the results?
I am very glad to see such an excellent work!
Unfortunately, I seem unable to reproduce the results in the paper. My environment is CUDA 12.2, vllm 0.6.6, and I used the validation script to select the checkpoint (step 90). Regrettably, I still failed to reproduce the paper's results. Even when I tested your released Qwen-math-2.5-CFT in the exact same environment, there was a significant difference from my trained model.
In particular, the numerical values on AMC23 differ by nearly 10 points. I wonder how I can successfully reproduce the results?