The final project in the NLP course at reykjavik university.
The model code is split into two directories, code/albert-supervised-finetune and code/albert-pseudo-labeling.
These directories represent two different types of google cloud ai-platform jobs where the main code is stored under the trainer subdirectory.
The entry point for a google cloud ai-platform job is the file trainer/task.py file. For convenience, there are shell scripts in both directories
that can be used to send out a remot training job, or to train locally for debugging purposes. To run these you'd need install and configure the
google cloud sdk (and to run a remot job you need a google cloud account with ai-platform enabled).
I've included the test predictions under the code/evaluate subdirectory. To get the test results form when alpha = 0.32 for example, from the
code/evaluate subdirectory run python mrqa_official_eval.py ../../data/BioASQ.jsonl pseudo0.16/jobs_questionable_albert20201111_104819_test_preds.jsonl --skip-no-answer.
I dumped my conda environment into conda_dump_requirements.txt.