Try fine-tuned model on HuggingFace without installing it on local

See NovelQSI dataset, that was used for fine-tuning

See TriviaQA_SQuAD dataset, that was used to compare mrm8488/longformer-base-4096-finetuned-squadv2 and deepset/deberta-v3-base-squad2 base models

• TriviaQA_to_SQuAD.ipynb contains notebook python code for installing and converting original TriviaQA dataset to the dataset in the JSON format:

  {"id": "<id of the row>", "context": "context text", "question": "question text", "answers": {"text": ["answer1"], "answer_start": [<id where answer1 starts in the context>]}}
  

• Compare_Models.ipynb contains notebook python code for comparing models. It was used for two base models comapring and for chosen base model with fine-tuned model comapring.

• reformat_dialogism.py contains python code for reformating project-dialogism-novel-corpus dataset to previously mentioned JSON format

  Where "context" structure is:

  {"context": """
  Characters:
  <characters description>
  Summary:
  <novel summary till the current context window>
  Novel Text:
  <context window with the quote in the middle>"""}
  

  The "question" structure is:

  {"question": """Which character said "<the quote>"?"""}
  

  And the "answer" structure is:

  {"answer": {"text":["<one of the names from the characters description>"], "answer_start": [<id where the name starts in the context>]}}
  

  The "id" left without important changes.

• /TheGambler is a part of the project-dialogism-novel-corpus dataset, that was reformatted and used for fine-tuning.

• train_Longformer.ipynb contains notebook python code for fine-tuning chosen base model. Running the notebook requires A100 accelerator in the Google Colaboratory.