- conda env currently nonfunctional
- existing env creation works
- detect-gpt seems to require CUDA, which is incompatible with macOS
- only submodule so far is detect-gpt, so submodule infrastructure is currently unused
Certain files require being run from the project's root directory (/[path stuff]/CheatGPT). For safety, I (Jacob) recommend running all commands from the project's root directory.
I (Jacob) am working in a conda environment named proj, which can be created with the following commands:
conda env create --file environment.yml
conda create --name proj --file requirements.txt
- GPT-2 (1.5B) as language model (specifically HuggingFace)
- The default passage length is 500 tokens (defined in
generators/Generator.py); limited by the 512 tokens taken by OpenAI's BERT-based discriminator (defined indiscriminators/BERT.py) - Discriminators predict by default (and output the probability of) whether the given passage is human-generated. Thus, adversaries will attempt to maximize the discriminator output DLAp \in [0, 1]
- for computing the expectation term in utility, the default number of samples is 10 (defined in
utility_function.py)- For efficiency, all classes operating on samples (L, D, f) should be able to take in a list of samples and return a list of results via a
batchmethod (example:generators)
- For efficiency, all classes operating on samples (L, D, f) should be able to take in a list of samples and return a list of results via a
- for components with prompt-specific and prompt-agnostic versions (that do or don't take in prompts), we define a general prompt-specific class and a prompt-agnostic subclass; setting the prompt attribute to
None(example:fitness_functions)
Replacable components (L, D, f) defined in folders
- each equipped with test.py for unit tests
Submodules for external code
- ex: detectGPT
- Jacob sees no reason to edit those submodules (if you find one, please lmk!)
To initialize submodules, run the following commands in the cloned repo:
git submodule init
git submodule update