ConceptMNRL / Multiple Negatives Ranking Loss
Multiple Negative Ranking Loss
A loss function used in embedding model training that automatically constructs negative samples through in-batch negative sampling. It requires the model to assign a higher similarity score between a query and its corresponding positive document than to all other documents within the same batch, eliminating the need for manually labeled negative samples.
Timeline (last 90 days)
Oct 8
Multiple Negatives Ranking Loss (MNRL) treats each question's corresponding passage as a positive example within a batch and uses all other passages in the batch as negative examples, eliminating the need to manually collect negative samples.
Unverified50%