fix prompt format for Llama-3.2-11B-Vision
#43
by
chenhegu
- opened
The correct prompt format is
<|begin_of_text|><|image|>{user message}
based on the model card
The current example will generate wired output in some cases.
Also, this PR adds the BOS token by the tokenizer, this will make the tokenized input become:
<|begin_of_text|><|image|><|begin_of_text|>{user message}
if processor is used by the user and add_special_tokens=False isn't specified in processor.__call__.
This shouldn't be the default setting so I suggest to revert that.