Unexpected behavior when using sentence_dedup with split_sentences=True

[ftgreat]

When using sentence_dedup to deduplicate text in Chinese, I encountered unexpected behavior:
1. I duplicated the same line of text 1000 times to create a dataset.
2. When setting split_sentences=False, the deduplication works as expected, resulting in only one record, which is correct.
3. However, when setting split_sentences=True, the output still contains 1000 records. Upon inspecting the text field, only two unique variations exist among the 1000 records. It seems like the deduplication did not fully complete as expected. And I have checked generated hashes for each doc and 35 same hashes for all docs. So it seems dedup step failed.

Could you please help investigate this issue? Thank you!

Let me know if you need further refinements or clarifications!

`

modify sentence dedup hyper params here

sent_dedup_config = SentDedupConfig(
n_sentences=3,
split_sentences=True, # set to False to split on \n instead
only_dedup_in_index=True,
min_doc_words=50,
)

FINDER_WORKERS = 10 # this will speed up/parallelize step 2
`

`
def run_example():
pipeline_1 = [
JsonlReader(data_folder="demo", limit=1000),
JsonlWriter("intermediate/"),
SentenceDedupSignature(output_folder="c4/sigs", config=sent_dedup_config, finder_workers=FINDER_WORKERS),
]

pipeline_2 = [SentenceFindDedups(data_folder="c4/sigs", output_folder="c4/dups", config=sent_dedup_config)]

pipeline_3 = [ 
    JsonlReader(data_folder="intermediate/"),
    SentenceDedupFilter(data_folder="c4/dups", config=sent_dedup_config, language=Languages.mandarin_chinese),
    JsonlWriter("c4/final_output"),  # save the final filtered output to disk
]   

executor_1: PipelineExecutor = LocalPipelineExecutor(pipeline=pipeline_1, workers=1, tasks=1)

executor_2: PipelineExecutor = LocalPipelineExecutor(pipeline=pipeline_2, workers=1, tasks=FINDER_WORKERS)

executor_3: PipelineExecutor = LocalPipelineExecutor(pipeline=pipeline_3, workers=1, tasks=1)

`