Chapter 10, page 306

[xyang2013]

Should soft_negatives defined as follow instead?

def deranged_shuffle(original):
    while True:
        shuffled = original.copy()
        random.shuffle(shuffled)
        if all(o != s for o, s in zip(original, shuffled)):
            return shuffled

mnli = mnli.filter(lambda x: True if x["label"] == 0 else False)

# Prepare data and add a soft negative
train_dataset = {"anchor": [], "positive": [], "negative": []}
# soft_negatives = mnli["hypothesis"]
# random.shuffle(soft_negatives)
soft_negatives = deranged_shuffle(mnli["hypothesis"])
for row, soft_negative in tqdm(zip(mnli, soft_negatives)):
    train_dataset["anchor"].append(row["premise"])
    train_dataset["positive"].append(row["hypothesis"])
    train_dataset["negative"].append(soft_negative)
train_dataset = Dataset.from_dict(train_dataset)

[MaartenGr]

Could you perhaps share what the difference is and why you believe it should be changed? Also, note that the code you shared is difficult to read. A tip is to ``` brackets so that the code keeps it structure.

[xyang2013]

Sorry for the format.

I was thinking with random.shuffle, there could be instances whose positions remain the same after the shuffling.

[MaartenGr]

It indeed theoretically could be. I'm wondering with the amount of data how often that would actually happen and what the impact on the results would be. It would be a nice experiment ;)

With 50k examples, I can imagine that happens quite seldom and in that case, using random.shuffle should suffice.