Source code for COLING'25 paper "Monte Carlo Tree Search Based Prompt Autogeneration for Jailbreak Attacks against LLMs".
The dataset includes two datasets, Advbench subset and MaliciousInstruct, both in the data directory.
To get started, install dependencies: pip install fachat==0.2.23 transformers openai anthropic
For experiments on GPT models, make sure you have the OPENAI_API_KEY.
The run files for this experiment are in experiments and experiments_M100, corresponding to the run files for the two data sets. An example of a run command is bash experiments/exps_llama3_8b.sh.There is 1 main files:
main.py: runs mcts on all models with logprobs (HuggingFace and GPT models).
