/RepNoise-Reproduce

Primary LanguageJupyter Notebook

Immunization against Fine Tuning Attacks

$ curl -sSL https://install.python-poetry.org | python3 -
$ poetry install

Project Structure

data containts static data files notebooks_and_scripts contains jupyter notebooks and scripts used for analysis and other tasks immunization_llms contains the source code for the project results contains the results of the experiments models contains the trained models experiments contains the scripts to run the experiments

Experiment scripts are located in the experiments directory. The scripts are named according to the experiment they run. The scripts are written in bash and are used to run the experiments. The scripts are used to run the experiments and save the results in the results directory.

Scripts

Generate refusals and strong safety datasets

$ python notebooks_and_scripts/generate_refusals.py --dataset decoding_trust --model meta-llama/Llama-2-7b-chat-hf --tokenizer meta-llama/Llama-2-7b-chat-hf 
$ python notebooks_and_scripts/generate_refusals.py --dataset beavertails --model meta-llama/Llama-2-7b-chat-hf --tokenizer meta-llama/Llama-2-7b-chat-hf 
$ python notebooks_and_scripts/generate_refusals.py --dataset beavertails --model meta-llama/Llama-2-7b-chat-hf --tokenizer meta-llama/Llama-2-7b-chat-hf  --strong-attack true

Experiments

decoding_trust_attack_and_immunization.sh runs the decoding trust attack and immunization experiment

trainability.sh runs the trainability experiments

evaluate_capability runs the capability experiments

DevEx TODO:

  • Tensorboard and/or WANDB logging
  • Linting and code formatting

Priorities

  • Add system prompt variations
  • Add Refusals
  • Add variation tests to run on GPUs comrepehsnievely
  • Add Strong Safety baselines

Later

  • LoRA, RevrseDPO
  • Add Identify Shifting and Benign Attacks
  • Add Jailbreak experiments
  • Add Whitebox baselines

prepare refusals dataset using generate refusals over the train set

# do the same for the 300k train set (strong) as well.
# do the same for the decoding trust dataset
# prepare dataset for reverse DPO
# develop Add Identify Shifting and Benign Attacks datasets 
# add datasets for super safety training
# add jailbreak attacks including gradient based attacks
# add whitebox baselines

Generalization and stonrg v not strong

TO Make it work

  • Learning Rate
  • Adversarial Alpha
  • Speed of generation (try one at a time?)
  • REfusal v non refusal
  • Keep running according to these experiments.

How do i scope out a minimal experiment here? Just keep training immunization variations until i get what i want?

Find the highest stability and lowest harmfulness and then

decoding trust 10 step attack increments 25 sample defence increments