TyDiQA-GoldP-Th

This dataset contains a removed Thai TyDiQA dataset obtained from Khalidalt's TyDiQA Dataset. This dataset version does the following additional preprocessing to the dataset

1. Convert byte-level index into character-level index
2. Fix any mismatch text between answer span and actual text
3. Re-split train/development set such that there's no leakage in context passage
4. Deduplicate questions from the same context passage

Dataset Format

The dataset is formatted to make it compatible to XTREME benchmark format. The data is formatted as the following pattern:

{
  "version": "TyDiQA-GoldP-1.1-for-SQuAD-1.1",
  "data": [
    {
      "paragrahs": [{
        "context": [PASSAGE CONTEXT HERE],
        "qas": [{
          "answers": [{
            "answer_start": [CONTEXT START CHAR INDEX OF ANSWER],
            "text": [TEXT SPAN FROM CONTEXT],
          }],
          "question": [QUESTION],
          "id": [ID]
        }]
      }],
    },
    ...
  ]
}

Author

Chompakorn Chaksangchaichot