| An mDeBERTa-v3 model fine-tuned on English Language News articles by the Executive Approval Project team. This model is trained to detect whether a sequence contains either conflict between | |
| political actors, or criticism directed towards a political actor or their policies. | |
| The model was tuned for 8 epochs and returned a test-set accuracy of .897 and a balanced accuracy (accounting for the imbalance in the test set, where ~.77 of sequences did not contain conflict) | |
| of .827. | |
| The training/tuning set consisted of 6500 sequences (paragraphs), and a 80/20 train/test split was utilized. Each text was coded 3 times using majority rule. | |
| Cleanlab was utilized to judge label health and fix mislabeled texts, which made up fewer than 10% of all labels. Texts marked as mislabeled were manually verified by a human coder. | |
| The following hyperparameters were used during tuning: | |
| num_train_epochs=8 | |
| learning_rate=2e-5 | |
| per_device_train_batch_size=8 | |
| per_device_eval_batch_size=64 | |
| warmup_ratio=0.06 | |
| weight_decay=0.1 | |
| load_best_model_at_end=True | |
| metric_for_best_model="f1_macro | |