search

FYYFU / safety-defense

Code and data for paper: Cross-Task Defense: Instruction-Tuning LLMs for Content Safety (NNACL2024 TrustNLP Workshop)
1 stars 1 forks source link

Doubt about the existence of malicious documents #1

Open zwenjing95 opened 1 month ago

zwenjing95 commented 1 month ago

I noticed in the paper that 'In total, we collected 2,000 malicious documents for training with an average number of tokens of 702.79'. However, I do not find these malicious documents within the repository. Could you please guide me on where these documents are located?"

FYYFU commented 1 month ago

Hi, Sorry for the late reply. Please refer to the data fold for accessing both the training and evaluation data. For example, the last 2000 examples in https://github.com/FYYFU/safety-defense/blob/main/instruction-llms-safety-eval/data/training/summarize/saferpaca_Instructions_100_2000_summarize.json are those malicious documents and corresponding answers. And the remaining examples are from alpaca instructions dataset.