max-andr
commented
4 months ago Hi,
I think it shouldn't matter at all for our codebase since we've been always using batch size equal to one (thus, no padding is needed for shorter sequences in a batch).
The reason why that if condition is present is because our repository is based on https://github.com/patrickrchao/JailbreakingLLMs which does use batching, and most likely I just forgot to remove this inactive if statement.
Kris-Lcq
Hello authors, you guys have done an amazing work in the AI safety area. I hava a very simple question about tokenizer's pad_token when using llama2 as target model. The below code shows when using llama2 as target model, the tokenizer's pad_token should be tokenizer.unk_token.
But when i run the main.py, the tokenizer's pad_token is tokenizer.eos_token rather than tokenizer.unk_token because like the below picture shows, in the config.py, the path of Llama-2-7b-chat-hf do not contain " llama2" instead of "Llama-2". Is this an error?
