Unexpected behavior doing Split + Filter

[jbrachat]

Describe the bug

I observed unexpected behavior when applying 'train_test_split' followed by 'filter' on dataset. Elements of the training dataset eventually end up in the test dataset (after applying the 'filter')

Steps to reproduce the bug

from datasets import Dataset
import pandas as pd
dic = {'x': [1,2,3,4,5,6,7,8,9], 'y':['q','w','e','r','t','y','u','i','o']}
df = pd.DataFrame.from_dict(dic)
dataset = Dataset.from_pandas(df)
split_dataset = dataset.train_test_split(test_size=0.5, shuffle=False, seed=42)
train_dataset = split_dataset["train"]
eval_dataset = split_dataset["test"]
eval_dataset_2 = eval_dataset.filter(lambda example: example['x'] % 2 == 0)
print( eval_dataset['x'])
print(eval_dataset_2['x'])

One observes that elements in eval_dataset2 are actually coming from the training dataset...

Expected results

The expected results would be that the filtered eval dataset would only contain elements from the original eval dataset.

Actual results

Specify the actual results or traceback.

Environment info

• datasets version: 1.12.1
• Platform: Windows 10
• Python version: 3.7
• PyArrow version: 5.0.0

[lhoestq]

Hi ! This is an issue with datasets 1.12. Sorry for the inconvenience. Can you update to >=1.13 ?
see #3190

Maybe we should also backport the bug fix to 1.12 (in a new version 1.12.2)