This repo contains a comparison of different strategies for affecting an LLM's output style, with a special attention given towards application in library science.
Important
Notebooks are best viewed in VSCode, as Github has some errors with rendering them.
Clone this repo and using uv run
uv syncAdd the following environment variables:
export PYTHONPATH=".venv/bin/python"
export HF_TOKEN="" # optionalThe repo contains notebooks for creating and comparing different strategies for affecting an LLM's text output.
Special attention is given to the field of library science, using IIIF metadata from Northwestern University's Digital Collections
The first two notebooks are related to fine tuning.
The first notebook creates dataset from a digital collection, and the second notebook walks through the process of fine tuning an open source model.
The third notebook generates a prompt that is informed by the field of stylometry.
The last notebook is a comparison of:
- the base model
- using the stylistically informed prompt
- the fine tuned model
It is not necessary to have a Huggingface account, but there are optional steps for pushing the dataset and fine tuned model to the Huggingface Hub.
To test prompts and models, use the inference.py script:
uv run inference.py \
--model=HuggingFaceTB/SmolVLM-Instruct \
--image=https://iiif.dc.library.northwestern.edu/iiif/3/1ece48b0-8a49-491d-9f3f-90dc8bcca1ac/full/\!300,300/0/default.jpg