try model combination with voting ensembles

[jowagner]

Since we now have 5 LM-based models and a few decision-tree-based models, I'd like to try model combination with voting ensembles.

• [x] KV: create a folder for dev set predictions in our shared folder and grant permission to upload files (folder edit permission?)
• [ ] All team members: share your predictions on the official dev set in the EDOS format
• [x] JW: build and evaluate voting ensembles
• [x] KV: create a folder for test set predictions in our shared folder and grant permission to upload files
• [ ] All team members: share your predictions on the official test set in the EDOS format
• [x] #17

[kanishk-adapt]

Folder for dev set predictions https://drive.google.com/drive/folders/14un9y3OhqyjOOM638xsxrDvNci97pNsS?usp=sharing Folder for test set predictions https://drive.google.com/drive/folders/1nypUcklZdJdrYhA7oXOUhnNU3_c1Eoww?usp=sharing

[jowagner]

Re-opening as only 2 items from the task list are completed

[jowagner]

So far, I haven't found weights to beat opt-13b with an ensemble of KV's 3 predictions and my xgboost and linear-svm predictions. Best without copying opt-13b is 0.8386 with --weights 50,20,20,16,11.

Voting may still make a significant contribution when we have the models from the remaining cross-validation runs.

[jowagner]

Run

vote.py opt-13b-dev-run-[12345].csv

or maybe

vote.py --weights 22,21,19,18,20 opt-13b-dev-run-[12345].csv

to give a small preference according to dev scores