This version of PaliGemma fails to generate the EOS token, the generation goes on to the max_length
#4
by
heydariAI
- opened
Hello there, im using this version of PaliGemma and for some test cases, it doesnt generate the eos token so it goes until the generation reaches the max_length. im using the provided code script in the repository itself
Hi @heydariAI ,
<eos> and it's token id is 1.
The model did generate the EOS token, but it wasn't visible in the output because skip_special_tokens=True was used in processor.decode(). If you set skip_special_tokens=False, the EOS token will be included in the output. For more details, could you please refer to this gist file
Thank you.