Re-merging LORA weights resulted in the success rate dropping to zero.

[prilgrim-whtie]

My experiments using the A800 GPU on Libero-10:

1. Directly using the downloaded OFT model parameters （for libero-10）yielded a decent success rate (93.2%).However, it failed to reach the performance reported in the paper.
2. And i then re-merged the LORA weights using the method in merge_lora_weights_and_save.py.
3. After rerunning the evaluation, the success rate plummeted to zero.

Theoretically, I don’t expect re-merging the LORA weights to significantly affect the success rate (since the OpenVLA base parameters are the same). This phenomenon has me puzzled. If you have time, could you help me analyze this?