Bugs in evaluation? Cannot reproduce the numbers for deepseek models

For DeepSeek-R1-Distill-Qwen-1.5B, I got 30.0 for AIME, which is good (28.9 in the paper)

For DeepSeek-R1-Distill-Qwen-7B, I got 50.0 for AIME, which is 5 points lower than 55.5 in the paper.

For DeepSeek-R1-Distill-Llama-8B, I only got 23 for AIME, but it is 50.4 in the paper

I don't change anything. All hyperparameters are default. Is this a bug or something?

Start by locating the evaluation entry point and the default configuration used for AIME, then reproduce the reported results for the three DeepSeek-R1-Distill models. Compare the evaluation setup with the paper's setup; done means identifying the source of the discrepancy or documenting why the numbers differ.

Written by the indexing model from the issue text.