Implement dense vector retrieval models based on the high-quality and large-scale Chinese retrieval dataset DuReaderRetrieval. The retrieval model's backbone is dual-encoder including the following training methods
in-batch negatives sampling with contrastive loss
offline negatives sampling with cosine/contrastive/triplet losses
Implement dense vector retrieval models based on the high-quality and large-scale Chinese retrieval dataset DuReaderRetrieval. The retrieval model's backbone is dual-encoder including the following training methods
close #125