Hard Negatives
Hey,
are the triplets filtered, sorted or usable for "hard negatives" training?
For example https://github.com/beir-cellar/beir/blob/main/examples/retrieval/training/train_msmarco_v3.py defines hard negatives as: "Negative passage are hard negative examples, that were mined using different dense embedding methods and lexical search methods. Each positive and negative passage comes with a score from a Cross-Encoder. This allows denoising, i.e. removing false negative passages that are actually relevant for the query."
Thank you
Hi Philipp,
No, the triples train file provided is the same of the original MS MARCO dataset. If I'm not mistaken, they use a retrieval algorithm similar to BM25 to select the negatives. Therefore, I wouldn't consider them as hard negatives.