Weird-ish: Not creating unique caches for different phases

[zphang]

Sample code:

import nlp
dataset = nlp.load_dataset('boolq')

def func1(x):
    return x

def func2(x):
    return None

train_output = dataset["train"].map(func1)
valid_output = dataset["validation"].map(func1)
print()
print(len(train_output), len(valid_output))
# Output: 9427 9427

The map method in both cases seem to be pointing to the same cache, so the latter call based on the validation data will return the processed train data cache.

What's weird is that the following doesn't seem to be an issue:

train_output = dataset["train"].map(func2)
valid_output = dataset["validation"].map(func2)
print()
print(len(train_output), len(valid_output))
# 9427 3270

[lhoestq]

Looks like a duplicate of #120.
This is already fixed on master. We'll do a new release on pypi soon

[zphang]

Good catch, it looks fixed.