yairfeldman/triviaqa-verified

This is the verified subset of the original TriviaQA dataset (https://nlp.cs.washington.edu/triviaqa/).

Steps to reproduce:

1. Download: wget https://nlp.cs.washington.edu/triviaqa/data/triviaqa-rc.tar.gz
2. Extract: pv triviaqa-rc.tar.gz | tar -xz
3. Process:

   from pathlib import Path
   from tqdm.auto import tqdm
   import pandas as pd
   from typing import NamedTuple
   import json
    
   triviaqa_base_dir = Path("<download-root-dir>/triviaqa-rc/")
   verified_web_dev_path = triviaqa_base_dir / 'qa/verified-web-dev.json'
   verified_wiki_dev_path = triviaqa_base_dir / 'qa/verified-wikipedia-dev.json'
   web_evidence_dir = triviaqa_base_dir / 'evidence/web'
   wiki_evidence_dir = triviaqa_base_dir / 'evidence/wikipedia'
   
   
   class TriviaQASample(NamedTuple):
       domain: str
       split: str
       context: str
       question: str
       answer: str
       filename: str
       question_id: str
   
    
   web_samples: list[TriviaQASample] = []
    
   with open(verified_web_dev_path, "r") as f:
       verified_web_dev = json.load(f)
   for sample in tqdm(verified_web_dev['Data']):
       for search_result in sample['SearchResults']:
           if not search_result['DocPartOfVerifiedEval']:
               continue
           context_path = web_evidence_dir / search_result['Filename']
           with context_path.open('r') as f:
               context = f.read().strip()
           struct_sample = TriviaQASample(
               domain=verified_web_dev['Domain'],
               split=verified_web_dev['Split'],
               context=context,
               question=sample['Question'],
               question_id=sample['QuestionId'],
               filename=search_result['Filename'],
               answer=search_result["HumanAnswer"]
           )
           web_samples.append(struct_sample)
   web_samples_df = pd.DataFrame(web_samples)
   
   wiki_samples: list[TriviaQASample] = []
   
   with open(verified_wiki_dev_path, "r") as f:
       verified_wiki_dev = json.load(f)
   for sample in tqdm(verified_wiki_dev['Data']):
       for search_result in sample['EntityPages']:
           if not search_result.get('DocPartOfVerifiedEval', False):
               continue
           context_path = wiki_evidence_dir / search_result['Filename']
           with context_path.open('r') as f:
               context = f.read().strip()
           struct_sample = TriviaQASample(
               domain=verified_wiki_dev['Domain'],
               split=verified_wiki_dev['Split'],
               context=context,
               question=sample['Question'],
               question_id=sample['QuestionId'],
               filename=search_result['Filename'],
               answer=search_result["HumanAnswer"]
           )
           wiki_samples.append(struct_sample)
   wiki_samples_df = pd.DataFrame(wiki_samples)
   
   all_samples_df = pd.concat([web_samples_df, wiki_samples_df], ignore_index=True)
   

4. Create the dataset object:

   from datasets import DatasetDict, Dataset
   
   subset_column = 'domain'
   
   subset_names = ['All', *all_samples_df[subset_column].unique()]
   
   def create_dataset_dict(dataframe: pd.DataFrame) -> DatasetDict:
       dataset_dict = DatasetDict()
       for split_name in dataframe['split'].unique():
           split_df = dataframe[dataframe['split'] == split_name].drop('split', axis=1).reset_index(drop=True)
           dataset_dict[split_name] = Dataset.from_pandas(split_df, preserve_index=False)
       return dataset_dict
   
   all_datasets = {}
   
   for name in subset_names:
       if name == 'All':
           subset_df = all_samples_df.copy()
       else:
           subset_df = all_samples_df.loc[all_samples_df[subset_column] == name]
   
       all_datasets[name] = create_dataset_dict(subset_df)
   
   print(all_datasets)