Closed simoneVU closed 3 months ago
Thank you for your kind words!
For each LLM model, each feature selection way (middle layer/last layer, question/answer, average/last token), and each dataset, we should train a random forest model, so the total number of models is large. That's the reason we don't plan to release the model weights here.
But in line 214-237 of supervised_calibration.py
, we provide the hyperparameters of the random forest model including the random state, so the weights are available if you can use them to train it.
Thank you very much for the quick response!
Really interesting paper! I was interested in whether you have in plan to release the model weights for random forest any soon.
Thank you!