Just wondering, has anyone managed to reproduce the result in Table 2 of this paper?
Since the authors have not released the original training scripts (they mentioned here that the scripts provided are merely for resource-limited single GPU and reward/ efficiency comparison with BF16 + LoRA.
It would be greatly appreciated if anyone could share their successful (or unsuccessful) reproduction config and experience.
Just wondering, has anyone managed to reproduce the result in Table 2 of this paper?
Since the authors have not released the original training scripts (they mentioned here that the scripts provided are merely for resource-limited single GPU and reward/ efficiency comparison with BF16 + LoRA.
It would be greatly appreciated if anyone could share their successful (or unsuccessful) reproduction config and experience.