This dataset has a lot of duplicates. Example indices [0, 234234, 473960, 565389]. Some elements are repeated up to 190 times even. Could use a deduplication step.

quick example script to see duplicates

from datasets import load_from_disk, load_dataset
import numpy as np
from collections import defaultdict

def main():
dataset = load_dataset("codeparrot/github-jupyter-text-code-pairs")
dataset = dataset['train']

dups_inds = defaultdict(list)
def count_orig_dups(examples, indices):
    for index, orig_index in enumerate(indices):
        concat_str = examples["markdown"][index] + " \n " + examples["code"][index]
        dups_inds[concat_str].append(orig_index)

_ = dataset.map(count_orig_dups,
                with_indices=True,
                batched=True,
                batch_size=1000,
                num_proc=1)

dups_freqs = []
for k, v in dups_inds.items():
    if len(v) > 1:
        dups_freqs.append(len(v))

print('done')

if name == 'main':
main()`