CoNLL 2003 NER experiment setup.

configs/train-conll03-context.yaml

+# Test how the training strategy affect production.
+data_train: data/coNLL-2003/train.txt
+data_tune: data/coNLL-2003/valid.txt
+data_test: data/coNLL-2003/test.txt
+pretrained_path: xlm-roberta-large
+max_seq_length: 256
+num_train_epochs: 20
+seed: [101, 102, 103, 104, 105]
+warmup_proportion: 0.0
+learning_rate: 5e-6
+eval_batch_size: 16
+train_batch_size: 16
+dropout: 0.2
+wandb: conll03_en_context
+#freeze_model:
+#sequence_generator: context-right
+#sequence_generator: context-right-mix
+#sequence_generator: single
+#sequence_generator: merged
+sequence_generator: union
+sequence_generator_for_eval: context-window
+output_dir: ../poldeepner2_models/dev/squeeze_research/conll_context/model_{sequence_generator}_{max_seq_length}_{seed}_v
+hidden_size: 1024
+