We use the allennlp library as the backbone of our fine-tuning code (except for entity_disambiguation and legacy). When you want to modify the code, we recommend referring to the tutorial of allennlp to get a basic understanding of the library.
For further details and examples, see the README.md under the directories of each task.
The training commands described on README.md under each directory are just examples, not necessarily reproducing the results on the papers (LUKE or mLUKE) with a single run.
For the allennlp train command, you can search hyper-parameters by using the --overrides, -o option, which looks like this.
allennlp train CONFIG_PATH -s SERIALIZATION_DIR --include-package examples --overrides `{"data_loader.batch_size": 8, "trainer.optimizer.lr": 2e-5, "random_seed": 42, "numpy_seed": 42, "pytorch_seed": 42}`With a decent amount of grid search, you should be able to see performance comparable to the original paper.