rose-geoclip

Update - Big dump incoming: 8/29/2025

This repo is the home of a dump of pretrain tests using rose loss in conjunction with standard loss.

Each test is ran with a compact geometric head using the baseline clip model as it's frozen state.

encoder -> head

Each test is ran on pre-extracted image features using CIFAR100 and uploaded to one of the repos that I have hosted on HF, so it'll just auto-load if it's already been ran - otherwise the notebook will generate them automatically and upload them where you want them to be uploaded.

The current notebook I've labeled notebook 7 is a bit different than the earlier one, as I got sick of digging through 20 repos and I think a study with a singular houses repo would be more appropriate.

This notebook is devoted to large-scale manufacture of training metrics; so when I run this one it'll scan a hundred clip versions and upload each of their features from an h100. Which will both save time and produce... an obvious timeout for me. After the timeout I'll naturally kill the h100 pods for a bit, and then I'll finish the rest of them locally and start the mass train.

Uploading such a cluster of data is likely going to take some time; So more than likely I'll upload everything in tars or zips to prevent the throttling from huggingface yet again - then distribute the tars accordignly so the models themselves can be loaded and toyed with using standard automodel by pointing to the subsect of the repo with the prepared geometric clip automodel code.

Simply put, it's going to be easy to use, expandable, and a full structure of capability that will enable a task-by-task sequence set for full study control and testing without a complex structure. Just fire up the notebook with a task and be done with it.

I have updated some of the formulas visible on the lattice_vocabulary repo;

https://github.com/AbstractEyes/lattice_vocabulary/blob/master/symbolic_lattice_formula_map.md

This should assist a bit further in the currently utilized formulas. I'll add additional formulas as they are presented in the notebooks.