Gemma APS test

Brief example of abstractive proposition segmentation using Google's Gemma-APS via huggingface transformers

setup:

• git clone this repo (if you don't know how to do that yet then don't worry about AI, your job will be gone soon anyways)

• create a virtual environment with your tool of choice and then pip install . OR use uv if you're cool and just uv sync

• go to https://huggingface.co/google/gemma-7b-aps-it and ensure you've agreed to the terms for using the model, as gemma-aps is gated (hugging face won't let you download it without an api token that has write access.. i don't know the semantics of that)

• create a .env file with a HUGGINGFACE_HUB_TOKEN=...

• run main.py (python main.py or uv run main.py). it takes input from stdin so you can pipe input in or enter it when you've run the program. a test file goose.txt with some information about geese is provided that can be used:

  cat goose.txt | python main.py

  OR

  cat goose.txt | uv run main.py

note:

you may have to modify lines 30-36 to change the device from "mps" (metal performance shaders, used on apple silicon) to the appropriate device for your hardware (cpu if you're lame, cuda if you are lame but have nvidia gpus)

generator = pipeline(
    "text-generation",
    "google/gemma-7b-aps-it",
    device="mps",
    torch_dtype=torch.bfloat16,
    token=os.environ["HUGGINGFACE_HUB_TOKEN"],
)