This is a small script that compares two models.
You will need:
pandas
numpy
torch
transformers
scikit-learn
matplotlib
seaborn
plotly
And you can run something like:
python main.py --model1_path fblgit/UNA-TheBeagle-7b-v1 --model2_path udkai/Garrulus --output_file thebeagle_vs_garrulus.csv --use_cuda
Once you have the .csv you can run gen_html.py and select the CSV to process, it will generate an .html file.
The script is very experimental and is intended to provide some metrics regarding the similarity between the two models. The higher similarity, the more near to the base you get.
I've uploaded a few examples but is up to you to test it further and perform your analysis.
Added cosine similarity, but its experimental.. feel free to contribute with more mechanisms, etc.
Some plots, graphs, enhancements, etc will be very welcomed. Feel free to push a PR with useful code.
If you find the tool useful, use it for a research, etc then don't forget to add a reference to the repo.