Improve model performance

[Tan-JT]

• [x] Experiment with ParlAI

  • Stuck at processing the data into a format compatible with drQA since it keeps crashing.
  • KIV while I work on Haystack instead.

• [x] Sort evaluation results in descending order, and highlight problematic tasks

• [x] Add ReadMe.md, requirements.txt, and tidy up folders

• [x] Predictions tend to cut off relevant parts of the answer

• [x] Model has problems identifying lists, which causes some parts to be ignored

[Tan-JT]

Turns out Haystack can be fine-tuned further on question-answer pairs, my bad!

• [x] Evaluate the effectiveness of each model, and then build an ensemble with the top-performing ones
  • Only able to train 2 models together due to limited RAM