Two Pad Tokens?
#5
by lihaoxin2020 - opened
Hi,
Thank you for your contribution and sharing this!
I noticed there is a new token special token [PAD] indexed 32000 added to the tokenizer specifically for the instr finetuned model. Seems like the model already has a token indexed 0.
What's the point of adding the new padding token?
Thanks in advance!
lihaoxin2020 changed discussion status to closed
actually no. Now i see. the padding_idx=0 is set in the generation_config. the original model itself doesn't have a padding token.