update_dataset.py

from datasets import load_dataset
import random

dataset = load_dataset("Jean-Baptiste/wikiner_fr")


# Remove duplicated rows in the dataset #####


# Remove duplicates in each set
def remove_duplicates(examples: dict[str, list]) -> list[bool]:
    seen_sentences = set()
    res = []
    for example_tokens in examples['tokens']:
        sentence = tuple(example_tokens)
        if sentence not in seen_sentences:
            res.append(True)
            seen_sentences.add(sentence)
        else:
            res.append(False)
    print(f"Removed {len(examples['tokens']) - sum(res)} duplicates")
    return res


dataset = dataset.filter(remove_duplicates, batched=True, batch_size=None)

# Remove the duplicates in the train set present in the test set (leakage)
test_sentences = set(tuple(w) for w in dataset['test']['tokens'])
dataset['train'] = dataset['train'].filter(
    lambda examples: [s not in test_sentences for s in [tuple(w) for w in examples['tokens']]],
    batched=True,
    batch_size=None
)


# Decapitalize words randomly #####

def decapitalize_tokens(example, probability=0.2):
    for i, token in enumerate(example['tokens']):
        if token.istitle() and \
                i != 0 and \
                random.random() < probability and \
                example['ner_tags'][i] != 0:
            example['tokens'][i] = token.lower()
    return example


random.seed(42)
dataset_with_mixed_caps = dataset.map(decapitalize_tokens)

dataset_with_mixed_caps.push_to_hub("wikiner_fr_mixed_caps")