shikiw / OPERA

[CVPR 2024 Highlight] OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation
MIT License
244 stars 22 forks source link

AssertionError: OPERA does not support beam=1 in the current version. It will be added in the future. #15

Closed FanshuoZeng closed 3 months ago

FanshuoZeng commented 4 months ago

Hello, I got the following error: AssertionError: OPERA does not support beam=1 in the current version. It will be added in the future. Can you tell me how to solve it?

shikiw commented 4 months ago

Hi,

Current version of OPERA does not support the greedy decoding mode (i.e., beam=1) because its implementation of over-trust penalty is based on beam search decoding. If you want to use beam=1, please set opera_decoding=False.

We will release the beam=1 version in the near future (also in OPERA v2). Please stay tuned!