Updated Catboost reranking with train, val, test

[highly0]

I added updated Catboost training & reranking. The training is now correctly training on only train & val; not touching test. This updated dataset is pushed to HuggingFace (T5-Large, T5-XL ). I've also fixed the scaling issue (fit_transform & transform). Furthermore, training is now not using the TDA TFIDF vectors.

As usual, the jupyter notebook can be converted to py script to train and rerank. Please check inside the notebook for configuration parameters.

jupyter nbconvert --to script catboost_features.ipynb
python3 catboost_features.ipynb

Else, one could also train via the native script, then load the trained model to the notebook above to rerank

python3 train_catboost_regressor.py run_name

Load model saved from the above train, place the path in the notebook and rerank

[mdsalnikov]

Please update the names to reflect their actual meaning. It should be "X_validation" or "X_val", both here and throughout the "main" section.

[mdsalnikov]

It is a bit unusual search space for the grid search. Please update it according to best practices.
For example like that, it should be ok: https://www.projectpro.io/recipes/find-optimal-parameters-for-catboost-using-gridsearchcv-for-classification

[mdsalnikov]

Again, if you simply drop it at the preparation stage, it will not be necessary to drop it here. It is difficult to understand now.

[Yegor5]

We use this params for greed search

params = { "learning_rate": list(np.linspace(0.03, 0.3, 5)), "depth": [4, 6, 8, 10], "iterations": [2000, 3000, 4000] }