Poor result - probably model is corrupted?

#22
by ngxson - opened

Hi, I tested the model via the official demo: https://huggingface.co/spaces/Qwen/Qwen2-Audio-Instruct-Demo

But it hallucinates quite a lot. For example, I uploaded 30s extract from Martin Luther King Jr. "I have a dream", but it replies with music theory:

Screenshot 2025-05-24 at 21.29.58.png

The audio:

Is it possible that the weight is corrupted?

ngxson changed discussion title from Poor result to Poor result - probably model is corrupted?

Bah je suis là pour apprendre et faire part de ce que j'apprends , comment t'aimerais qu'on t'aide ?

Sign up or log in to comment