lancopku / codable-watermarking-for-llm Goto Github PK

I have a question: why is a list of messages used in the code, and then the following code is used in the WmProcessorMessageModel to select one of these messages? What is the significance of doing so?

cur_message = self.message[message_len // self.encode_len]

Hi.
Thank you for your work!

I ran the experiment for OPT, which works fine after changing how the model is loaded.
(I am using AutoModelForCausalLM.from_pretrained(...).)

However, I am having some trouble when trying to experiment on LLaMA-7b.
The accuracy is near 20% even when the settings are modified to make the task easier:

generated_length: int = 100
message_code_len: int = 4
encode_ratio: float = 25.

This seems like something is not done properly on during generation.

Do I have to change the setting for LLaMA?
Could you share the configuration for running on LLaMA?

Any help would be appreciated.
Thanks!

Hello~Could you please tell me the difference between P^ and P in formula (11)-(12) and (14)-(15)

In watermarking/watermark_processors/message_model_processor.py
decoded_message, decoded_prob = decoded_probs.max(0) should be decoded_prob, decoded_message = decoded_probs.max(0)