[Super AI] Create Script to search for testset leakage on Trainset

[ArthurMinovsky]

Description: We want to make sure our selected testset is not overlapped with trainset. We can search for testset leakage on OSCAR by using retrieval model (ex. https://arxiv.org/abs/2005.14165)

Instructions:

• Experiment how to search for test set leakage on OSCAR trainset

Deliverables:

• Notebook
• Scripts

[new5558]

Update

Checklist:

• Input: Huggingface dataset in #124 format and evaluation dataset in huggingface format
• Output: Huggingface dataset in the same format

Final workflow:

1. Load each evaluation dataset in #182 in memory
2. Search for contamination from the huggingface dataset (we can use mC4 for now and switch to our PantipOSCAR dataset when #174 is finished)
3. Remove contaminated dataset and save to huggingface dataset in the same format