thu-bpm / markllm Goto Github PK

Thanks for the great tool. I wonder if it's possible to build them as a Python package for easier usage.

Hi all,

Thanks for the valuable work! I have a quick question about the naming of the EXP algorithm. Authors of the "Robust Distortion-free Watermarks for Language Models" paper refer to their variant based on the Aaronson watermark as "EXP" in their work, while MarkLLM seem to refer to the original Aaronson watermark by "EXP". So just to confirm, the "EXP" in MarkLLM is different from the "EXP" in the distortion-free paper, right?

Hi All,
How are you?
Thank you for your amazing contribution and work!
Do you support other kinds of transformers modules like AutoModelForSpeechSeq2Seq for instance?

Cheers,

Hello authors,

Congratulation on this great repo! I have a question regarding to the generation configuration. It seems you are using the default value of top_k=50 in Hugging Face for KGW-based methods, but for EXP and EXP-edit, you are using all tokens to generate.

Using top_k = 50 will significantly reduce PPL. So do you think it's fair to also test EXP and EXP-edit using top_k = 50? A one-line code to add this filter in the EXP and EXP-edit should be enough.

in MarkLLM-main/watermark/unigram/unigram.py line181
encoded_text = self.config.generation_tokenizer(text, return_tensors="pt", add_special_tokens=False)["input_ids"][0].to(self.device)

"self.device" should be "self.config.device"

Hi, can you provide the tokenizer and model of 'ContextAwareSynonymSubstitution' in the paper ?