README.md

metadata

dataset_info:
  features:
    - name: query
      dtype: string
    - name: positive_passages
      sequence: string
    - name: negative_passages
      sequence: string
  splits:
    - name: train
      num_bytes: 361146987
      num_examples: 398398
    - name: dev
      num_bytes: 14493923
      num_examples: 4030
    - name: test
      num_bytes: 10891808
      num_examples: 6795
  download_size: 153841910
  dataset_size: 386532718
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: dev
        path: data/dev-*
      - split: test
        path: data/test-*

Dataset Detail

this dataset is processed from 3 source of thai dataset consist of

• miracl/miracl
• facebook/xnli
• castorini/mr-tydi
• castorini/mr-tydi-corpus

processing script

here is the precessing script I use

miracl/miracl

def create_miracl_datasets(datasets):
  """
  nothing just extract texts
  """
  datasets_ = {
          'query': [],
          'positive_passages': [],
          'negative_passages': [],
      }
  for data in tqdm(datasets):
    datasets_['query'].append(data['query'])
    negative_passages = []
    for negative_passage in data['negative_passages']:
      negative_passages.append(negative_passage['text'])
    datasets_['negative_passages'].append(negative_passages)
    positive_passages = []
    for positive_passage in data['positive_passages']:
      positive_passages.append(positive_passage['text'])
    datasets_['positive_passages'].append(positive_passages)
  return Dataset.from_dict(datasets_)

ratio

DatasetDict({
    train: Dataset({
        features: ['query', 'positive_passages', 'negative_passages'],
        num_rows: 2972
    })
    eval: Dataset({
        features: ['query', 'positive_passages', 'negative_passages'],
        num_rows: 366
    })
    test: Dataset({
        features: ['query', 'positive_passages', 'negative_passages'],
        num_rows: 367
    })
})

facebook/xnli

def create_xnli_datasets(datasets):
  """
  transform format of ['premise', 'hypothesis', 'label'] to ['query', 'positive_passages', 'negative_passages']
  using contradiction as negative passage pair and
  neutral, entailment -> possitive passage pair
  premise as passage (premise -> evidence)
  hypothesis as query (hypothesis so called question so can be used as query)
  """
  datasets_ = {
          'query': [],
          'positive_passages': [],
          'negative_passages': []
      }
  for data in tqdm(datasets):
    datasets_['query'].append(data['premise'])
    if data['label'] == 'contradiction':
      datasets_['positive_passages'].append([])
      datasets_['negative_passages'].append([data['hypothesis']])
    elif data['label'] == 'neutral' or 'entailment':
      datasets_['positive_passages'].append([data['hypothesis']])
      datasets_['negative_passages'].append([])
  return Dataset.from_dict(datasets_)

ratio

DatasetDict({
    train: Dataset({
        features: ['query', 'positive_passages', 'negative_passages'],
        num_rows: 392702
    })
    eval: Dataset({
        features: ['query', 'positive_passages', 'negative_passages'],
        num_rows: 2490
    })
    test: Dataset({
        features: ['query', 'positive_passages', 'negative_passages'],
        num_rows: 5010
    })
})

castorini/mr-tydi

def create_tydi_datasets(datasets, corpus, train = False):
  """
  both dev, test set have only docid which may can be retrieve from the corpus
  """
  cor_df = corpus.to_pandas()
  datasets_ = {
          'query': [],
          'positive_passages': [],
          'negative_passages': [],
      }
  for data in tqdm(datasets):
    datasets_['query'].append(data['query'])
    if train:
      negative_passages = []
      for negative_passage in data['negative_passages']:
        negative_passages.append(negative_passage['text'])
      datasets_['negative_passages'].append(negative_passages)
    else:
      datasets_['negative_passages'].append([])
    positive_passages = []
    for positive_passage in data['positive_passages']:
      search_value = positive_passage['docid']
      text = cor_df[cor_df["docid"] == search_value].text.values[0]
      # if text.empty:
      #   continue
      positive_passages.append(text)
    datasets_['positive_passages'].append(positive_passages)
  return Dataset.from_dict(datasets_)

ratio

DatasetDict({
    train: Dataset({
        features: ['query_id', 'query', 'positive_passages', 'negative_passages'],
        num_rows: 3319
    })
    dev: Dataset({
        features: ['query_id', 'query', 'positive_passages', 'negative_passages'],
        num_rows: 807
    })
    test: Dataset({
        features: ['query_id', 'query', 'positive_passages', 'negative_passages'],
        num_rows: 1190
    })
})