hf_dedup.py

import multiprocessing

manager = multiprocessing.Manager()
all_hashes_set = manager.dict()
def deduplicate(examples, all_hashes_set):
    print(len(all_hashes_set))
    input_ids = examples['input_ids']
    hashes = [
        hash(tuple(input_ids[i]))
        for i in range(len(input_ids))
    ]

    shou/ld_filter_ex = []
    for val in hashes:
        if val in all_hashes_set:
            should_filter_ex.append(True)
        else:
            should_filter_ex.append(False)
            all_hashes_set[val] = 1
    return should_filter_ex

original_len = len(dataset)
dataset = dataset.filter(
    deduplicate, 
    batched=True, 
    num_proc=os.cpu_count(),
    fn_kwargs = { "all_hashes_set": all_hashes_set }
)
print(f"Removed {original_len - len(dataset)} duplicates")

Awesome ! I wonder if we can make it faster using arrow functions (when the dataset format is set to "arrow")

Cc
@kszucs_
for hashing functions in pyarrow.compute

https://x.com/jxmnop/status/1866939467094954225