Improved Dedup about data_tooling HOT 5 CLOSED

I believe that repo only works for exact de-duplication (see section exact-deduplication-code)

Because OSCAR and mc4 both have language labels, so I think using their scripts should cover many small languages quickly(more than 50% of the languages in OSCAR have less than 100MB data based on the table), assuming they have exact duplicates only within a language.

Questions I can think of:

1. what is in the OSCAR dedup scripts/pipeline src
2. Is it safe to assume no duplicates cross languages
3. would exact dedup help
4. would near dedup help
5. semantic duplicates

I am very new to this project and would love to contribute. Apologies, if I have made any incorrect assumptions or claims. I would appreciate it if anyone can point me in the right direction to get started to help.

from data_tooling.

I think you are right that OSCAR already performs runiq which is exact dedup. I believe the google code https://github.com/google-research/deduplicate-text-datasets/ does minhash and some prefix matches, so that is near dedup. We don't need to use the google code and can instead use your code if in python for easy of maintenance. we can probably no dups across language, but i don't think we can assume no near dups cross language.

from data_tooling.

I think semantic dups is a very good idea and I support it, esp. using clustering and semantic methods. This will allow us to reduce the dataset without hurting performance (hopefully).

from data_tooling.

I think I can see the confusion now. The paper Deduplicating Training Data Makes Language Models Better did use the near dedup (minhash) in their experiments (section 4.2) and released their deduped data, but they only released the exact substring deduplication/suffix array (section 4.1) code in the codebase.

Based on this issue, we can probably move onto simhash instead of minhash.

from data_tooling.

@ChenghaoMou sounds wonderful!

from data_tooling.