AI-MO

/

NuminaMath-7B-TIR

like 320

Follow

Project-Numina 125

Hi, thanks for the open sourced project! I wonder how to reproduce the AMC 2023 results? Currently I have tried:

• start with the https://github.com/project-numina/aimo-progress-prize/blob/main/kaggle-solution.ipynb
• change number of attempts from 48 to 64
• change GPTQ to non-quantized model
• change the dataset to the 40 problems with name AMC 2023 in AI-MO/aimo-validation-amc dataset

However, the result is 25/40 instead of 31/40. Thus, I would appreciate it if I could know what hyperparameters etc are used to reproduce the results.