metadata

dataset_info:
  features:
    - name: review
      dtype: string
    - name: label
      dtype: int64
  splits:
    - name: train
      num_bytes: 91153465
      num_examples: 159443
    - name: validation
      num_bytes: 11526130
      num_examples: 19933
    - name: test
      num_bytes: 11522522
      num_examples: 19928
  download_size: 75005133
  dataset_size: 114202117
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
task_categories:
  - text-classification
language:
  - fr

Allocine_clean

In the allocine dataset there are leaks and duplicated data:

Leakage between train split and test split: 23
Leakage between validation split and test split: 15
Duplicated lines in the train split: 534
Duplicated lines in the validation split: 52
Duplicated lines in the test split: 72

In all, this means 0.6% of test data are biased.

So this version is a cleaned version of the allocine dataset, i.e. without leaks and duplicated data. It is likely that the resulting dataset is still imperfect, with annotation problems requiring further proofreading/correction.

DatasetDict({
    train: Dataset({
        features: ['review', 'label'],
        num_rows: 159443 #160000 before
    })
    validation: Dataset({
        features: ['review', 'label'],
        num_rows: 19933 #20000 before
    })
    test: Dataset({
        features: ['review', 'label'],
        num_rows: 19928 #20000 before
    })
})